WO2019237870A1 - 目标匹配方法及装置、电子设备和存储介质 - Google Patents

目标匹配方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2019237870A1
WO2019237870A1 PCT/CN2019/086670 CN2019086670W WO2019237870A1 WO 2019237870 A1 WO2019237870 A1 WO 2019237870A1 CN 2019086670 W CN2019086670 W CN 2019086670W WO 2019237870 A1 WO2019237870 A1 WO 2019237870A1
Authority
WO
WIPO (PCT)
Prior art keywords
image sequence
feature vector
frame
query
candidate
Prior art date
Application number
PCT/CN2019/086670
Other languages
English (en)
French (fr)
Inventor
张瑞茂
孙红斌
罗平
葛玉莹
任宽泽
林倞
王晓刚
Original Assignee
深圳市商汤科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市商汤科技有限公司 filed Critical 深圳市商汤科技有限公司
Priority to KR1020207007917A priority Critical patent/KR20200042513A/ko
Priority to SG11202003581YA priority patent/SG11202003581YA/en
Priority to JP2020515878A priority patent/JP6883710B2/ja
Publication of WO2019237870A1 publication Critical patent/WO2019237870A1/zh
Priority to US16/841,723 priority patent/US11222231B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
    • G06F16/73—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying
    • G06F16/532—Query formulation, e.g. graphical querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/56—Information retrieval; Database structures therefor; File system structures therefor of still image data having vectorial format
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
    • G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/761—Proximity, similarity or dissimilarity measures

Definitions

  • the present disclosure relates to the field of computer vision technology, and in particular, to a method and device for object matching, an electronic device, and a storage medium.
  • Target matching refers to returning videos or images in the database that have the same target as the query video or query image.
  • Target matching technology is widely used in security monitoring systems at airports, stations, campuses, and shopping malls. In related technologies, the accuracy of target matching is low.
  • the present disclosure proposes a target matching technical solution.
  • a target matching method including:
  • the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence are extracted separately, where the query image sequence includes the target to be matched;
  • a cooperative expression feature vector of the query image sequence based on a feature vector of each frame in the query image sequence and a self-expression feature vector of the candidate image sequence, and based on a feature of each frame in the candidate image sequence A vector and a self-expressing feature vector of the query image sequence, and determining a cooperative expression feature vector of the candidate image sequence;
  • Determining the query image based on the self-expression feature vector of the query image sequence, the co-expression feature vector of the query image sequence, the self-expression feature vector of the candidate image sequence, and the co-expression feature vector of the candidate image sequence A similarity feature vector between the sequence and the candidate image sequence;
  • a matching result between the query image sequence and the candidate image sequence is determined.
  • extracting the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence separately include:
  • the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence are extracted by the first sub-neural network.
  • the method further includes:
  • the dimensionality reduction processing is performed on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence through the first fully connected layer of the first sub-neural network to obtain the query image sequence.
  • the self-expressing feature vector of the query image sequence and the The self-expression feature vectors of the candidate image sequence include:
  • the feature vector of each frame in the candidate image sequence and the first dimension-reduced feature vector of each frame in the candidate image sequence are input into a second sub-neural network to determine the self-expressing feature vector of the candidate image sequence.
  • a feature vector of each frame in the query image sequence and a first dimensionality reduction feature vector of each frame in the query image sequence are input to a second sub-neural network to determine the Query the self-expressing feature vector of the image sequence, including:
  • Determining the query image based on the second dimensionality reduction feature vector of each frame in the query image sequence, the overall feature vector of the query image sequence, and the first dimensionality reduction feature vector of each frame in the query image sequence The self-expressing feature vector of the sequence.
  • the feature vector of each frame in the candidate image sequence and the first dimension-reduced feature vector of each frame in the candidate image sequence are input to a second sub-neural network to obtain the Self-expressing feature vectors of candidate image sequences, including:
  • Determining the candidate image based on a second dimensionality reduction feature vector of each frame in the candidate image sequence, an overall feature vector of the candidate image sequence, and a first dimensionality reduction feature vector of each frame in the candidate image sequence The self-expressing feature vector of the sequence.
  • Dimensional feature vector, determining the self-expressing feature vector of the query image sequence including:
  • the correlation between the second dimensionality-reduced feature vector of each frame in the query image sequence and the overall feature vector of the query image sequence is calculated by a parameterless correlation function to obtain the first correlation of each frame in the query image sequence Weights;
  • determining the self-expression feature vector of the candidate image sequence including:
  • the first dimensionality reduction feature vector of each frame in the candidate image sequence is weighted to obtain the self-expression feature vector of the candidate image sequence.
  • the first correlation weight includes a first normalized correlation weight
  • the first normalized correlation weight is obtained by performing a normalization process on the first correlation weight.
  • a collaborative expression feature vector of the query image sequence is determined based on a feature vector of each frame in the query image sequence and a self-expression feature vector of the candidate image sequence, and based on the The feature vector of each frame in the candidate image sequence and the self-expression feature vector of the query image sequence, and determining the collaborative expression feature vector of the candidate image sequence include:
  • the feature vector of each frame in the query image sequence, the first dimensionality reduction feature vector of each frame in the query image sequence, and the self-expressing feature vector of the candidate image sequence are input into a third sub-neural network to obtain A collaborative expression feature vector of the query image sequence;
  • the feature vector of each frame in the candidate image sequence, the first dimensionality reduction feature vector of each frame in the candidate image sequence, and the self-expressing feature vector of the query image sequence are input into a third sub-neural network to obtain A collaborative expression feature vector of the candidate image sequence.
  • the feature vector of each frame in the query image sequence, the first dimensionality reduction feature vector of each frame in the query image sequence, and the self-expressing feature vector of the candidate image sequence Input to the third sub-neural network to obtain the collaborative expression feature vector of the query image sequence, including:
  • the feature vector of each frame in the candidate image sequence, the first dimensionality reduction feature vector of each frame in the candidate image sequence, and the self-expressing feature vector of the query image sequence are input into a third sub-neural network to obtain
  • the collaborative expression feature vector of the candidate image sequence includes:
  • the self-expression feature vector of the candidate image sequence Reducing the dimensionality feature vector to obtain the collaborative expression feature vector of the query image sequence, including:
  • the correlation between the third dimensionality-reduced feature vector of each frame in the query image sequence and the self-expression feature vector of the candidate image sequence is calculated through a parameterless correlation function to obtain the second of each frame in the query image sequence. Relevant weights;
  • the self-expression feature vector of the query image sequence and the first of each frame in the candidate image sequence Reducing the dimensionality feature vector to obtain the collaborative expression feature vector of the candidate image sequence, including:
  • the second correlation weight includes a second normalized correlation weight
  • the second normalized correlation weight is obtained by performing a normalization process on the second correlation weight.
  • a feature vector to obtain a similarity feature vector between the query image sequence and the candidate image sequence including:
  • a similarity feature vector of the query image sequence and the candidate image sequence is obtained.
  • obtaining a similarity feature vector of the query image sequence and the candidate image sequence based on the first difference vector and the second difference vector includes:
  • determining a matching result between the query image sequence and the candidate image sequence based on the similarity feature vector includes:
  • a matching result of the query image sequence and the candidate image sequence is determined.
  • the method further includes:
  • the same pair of labeled data and a binary cross-entropy loss function are used to optimize network parameters.
  • the method before extracting a feature vector of each frame in the query image sequence, the method further includes:
  • the method further includes:
  • a matching result between the query video and the candidate video is determined based on a matching result between the query image sequence of the query video and the candidate image sequence of the candidate video.
  • segmenting the query video into multiple query image sequences includes:
  • the query video is divided into a plurality of query image sequences according to a preset sequence length and a preset step length, wherein the length of the query image sequence is equal to the preset sequence length, and adjacent query image sequences overlap
  • the number of images is equal to a difference between the preset sequence length and the preset step size
  • Segment candidate videos into multiple candidate image sequences including:
  • the candidate video is divided into multiple candidate image sequences according to a preset sequence length and a preset step length, where the length of the candidate image sequence is equal to the length of the preset sequence, and adjacent candidate image sequences overlap
  • the number of images is equal to the difference between the preset sequence length and the preset step size.
  • determining a matching result between the query video and the candidate video based on a matching result between the query image sequence of the query video and the candidate image sequence of the candidate video includes:
  • N is a positive integer
  • a matching result of the query video and the candidate video is determined.
  • a target matching device including:
  • An extraction module for extracting a feature vector of each frame in a query image sequence and a feature vector of each frame in a candidate image sequence, wherein the query image sequence includes a target to be matched;
  • a first determining module configured to determine a self-expressing feature vector of the query image sequence and the candidate based on a feature vector of each frame in the query image sequence and a feature vector of each frame in the candidate image sequence; Self-expressing feature vectors of image sequences;
  • a second determining module configured to determine a collaborative expression feature vector of the query image sequence based on a feature vector of each frame in the query image sequence and a self-expression feature vector of the candidate image sequence, and based on the candidate image A feature vector of each frame in the sequence and a self-expression feature vector of the query image sequence, determining a cooperative expression feature vector of the candidate image sequence;
  • a third determining module configured to be based on the self-expression feature vector of the query image sequence, the co-expression feature vector of the query image sequence, the self-expression feature vector of the candidate image sequence, and the co-expression feature of the candidate image sequence A vector to determine a similarity feature vector between the query image sequence and the candidate image sequence;
  • a fourth determining module is configured to determine a matching result between the query image sequence and the candidate image sequence based on the similarity feature vector.
  • the extraction module is configured to:
  • the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence are extracted by the first sub-neural network.
  • the apparatus further includes:
  • a dimensionality reduction module configured to perform dimensionality reduction processing on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence through the first fully connected layer of the first sub-neural network to obtain A first dimensionality reduction feature vector of each frame in the query image sequence and a first dimensionality reduction feature vector of each frame in the candidate image sequence.
  • the first determining module includes:
  • a first determining submodule configured to input a feature vector of each frame in the query image sequence and a first dimensionality reduction feature vector of each frame in the query image sequence into a second sub-neural network to determine the query Self-expressing feature vectors of image sequences;
  • a second determining submodule configured to input a feature vector of each frame in the candidate image sequence and a first dimensionality reduction feature vector of each frame in the candidate image sequence into a second sub-neural network to determine the candidate The self-expressing feature vector of the image sequence.
  • the first determining submodule includes:
  • a first dimensionality reduction unit configured to perform dimensionality reduction processing on a feature vector of each frame in the query image sequence through a second fully connected layer of the second sub-neural network to obtain each frame in the query image sequence
  • the second dimensionality reduction feature vector
  • a first average pooling unit configured to subject the second dimensionality reduction feature vector of each frame in the query image sequence to a time dimension average pooling process to obtain an overall feature vector of the query image sequence
  • a first determining unit configured to be based on a second dimensionality reduction feature vector of each frame in the query image sequence, an overall feature vector of the query image sequence, and a first dimensionality reduction feature of each frame in the query image sequence Vector to determine the self-expressing feature vector of the query image sequence.
  • the second determining submodule includes:
  • a second dimension reduction unit configured to perform dimension reduction processing on a feature vector of each frame in the candidate image sequence through a second fully connected layer of the second sub-neural network to obtain each frame in the candidate image sequence
  • the second dimensionality reduction feature vector
  • a second average pooling unit configured to subject the second dimensionality reduction feature vector of each frame in the candidate image sequence to an average pooling process in the time dimension to obtain the overall feature vector of the candidate image sequence;
  • a second determining unit configured to be based on a second dimensionality reduction feature vector of each frame in the candidate image sequence, an overall feature vector of the candidate image sequence, and a first dimensionality reduction feature of each frame in the candidate image sequence Vector to determine the self-expression feature vector of the candidate image sequence.
  • the first determining unit includes:
  • a first calculation subunit configured to calculate a correlation between a second dimensionality-reduced feature vector of each frame in the query image sequence and an overall feature vector of the query image sequence through a parameterless correlation function to obtain the query image sequence
  • the first correlation weight of each frame in the frame
  • a first weighting subunit for weighting a first dimensionality reduction feature vector of each frame in the query image sequence based on a first correlation weight of each frame in the query image sequence to obtain the query image sequence Self-expressing feature vector.
  • the second determining unit includes:
  • a second calculation subunit configured to calculate a correlation between a second dimensionality reduction feature vector of each frame in the candidate image sequence and an overall feature vector of the candidate image sequence by using a parameterless correlation function to obtain the candidate image sequence
  • the first correlation weight of each frame in the frame
  • a second weighting subunit configured to weight the first dimensionality reduction feature vector of each frame in the candidate image sequence based on the first correlation weight of each frame in the candidate image sequence to obtain the candidate image sequence Self-expressing feature vector.
  • the first correlation weight includes a first normalized correlation weight
  • the first normalized correlation weight is obtained by performing a normalization process on the first correlation weight.
  • the second determining module includes:
  • a third determining submodule configured to input a feature vector of each frame in the query image sequence, a first dimensionality reduction feature vector of each frame in the query image sequence, and a self-expressing feature vector of the candidate image sequence Obtaining a collaborative expression feature vector of the query image sequence in a third sub-neural network;
  • a fourth determining submodule configured to input a feature vector of each frame in the candidate image sequence, a first dimensionality reduction feature vector of each frame in the candidate image sequence, and a self-expressing feature vector of the query image sequence
  • a collaborative expression feature vector of the candidate image sequence is obtained.
  • the third determining submodule includes:
  • a third dimension reduction unit configured to perform dimension reduction processing on a feature vector of each frame in the query image sequence through a third fully connected layer of the third sub-neural network to obtain each frame in the query image sequence
  • the third dimensionality reduction feature vector
  • a third determining unit configured to be based on a third dimension reduction feature vector of each frame in the query image sequence, a self-expression feature vector of the candidate image sequence, and a first dimension reduction of each frame in the query image sequence A feature vector to obtain a cooperatively expressed feature vector of the query image sequence;
  • the fourth determining sub-module includes:
  • a fourth dimension reduction unit configured to perform dimension reduction processing on a feature vector of each frame in the candidate image sequence through a third fully connected layer of the third sub-neural network to obtain each frame in the candidate image sequence
  • a fourth determining unit configured to be based on a third dimensionality reduction feature vector of each frame in the candidate image sequence, a self-expression feature vector of the query image sequence, and a first dimensionality reduction of each frame in the candidate image sequence A feature vector to obtain a collaborative expression feature vector of the candidate image sequence.
  • the third determining unit includes:
  • a third calculation subunit configured to calculate a correlation between a third dimension-reduced feature vector of each frame in the query image sequence and a self-expression feature vector of the candidate image sequence by using a parameterless correlation function to obtain the query image
  • the second correlation weight of each frame in the sequence
  • a third weighting subunit for weighting the first dimensionality reduction feature vector of each frame in the query image sequence based on the second correlation weight of each frame in the query image sequence to obtain the query image sequence Co-expression feature vector.
  • the fourth determining unit includes:
  • a fourth calculation subunit configured to calculate a correlation between a third dimensionality reduction feature vector of each frame in the candidate image sequence and a self-expression feature vector of the query image sequence by using a parameterless correlation function to obtain the candidate image
  • the second correlation weight of each frame in the sequence
  • a fourth weighting subunit configured to weight the first dimensionality reduction feature vector of each frame in the candidate image sequence based on the second correlation weight of each frame in the candidate image sequence to obtain the candidate image sequence Co-expression feature vector.
  • the second correlation weight includes a second normalized correlation weight
  • the second normalized correlation weight is obtained by performing a normalization process on the second correlation weight.
  • the third determining module includes:
  • a first calculation submodule configured to calculate a difference between a self-expression feature vector of the query image sequence and a collaborative expression feature vector of the candidate image sequence to obtain a first difference vector
  • a second calculation submodule configured to calculate a difference between a self-expression feature vector of the candidate image sequence and a co-expression feature vector of the query image sequence to obtain a second difference vector
  • a fifth determination submodule is configured to obtain a similarity feature vector of the query image sequence and the candidate image sequence based on the first difference vector and the second difference vector.
  • the fifth determining submodule includes:
  • a first calculation unit configured to calculate a sum of the first difference vector and the second difference vector to obtain a similarity feature vector between the query image sequence and the candidate image sequence;
  • a second calculation unit is configured to calculate a product of the first difference vector and elements of corresponding bits of the second difference vector to obtain a similarity feature vector of the query image sequence and the candidate image sequence.
  • the fourth determining module includes:
  • a sixth determining submodule configured to input a similarity feature vector of the query image sequence and the candidate image sequence into a fourth fully connected layer to obtain a matching score between the query image sequence and the candidate image sequence;
  • a seventh determination submodule is configured to determine a matching result between the query image sequence and the candidate image sequence based on a matching score of the query image sequence and the candidate image sequence.
  • the apparatus further includes:
  • An optimization module is used to optimize network parameters based on matching scores of the query image sequence and the candidate image sequence, using the same pair of labeled data and a binary cross-entropy loss function.
  • the apparatus further includes:
  • the first sub-module is used to divide the query video into multiple query image sequences
  • a second segmentation module configured to segment a candidate video into multiple candidate image sequences
  • a fifth determining module is configured to determine a matching result between the query video and the candidate video based on a matching result between the query image sequence of the query video and the candidate image sequence of the candidate video.
  • the first segmentation module is configured to:
  • the query video is divided into a plurality of query image sequences according to a preset sequence length and a preset step length, wherein the length of the query image sequence is equal to the preset sequence length, and adjacent query image sequences overlap
  • the number of images is equal to a difference between the preset sequence length and the preset step size
  • the second segmentation module is configured to:
  • the candidate video is divided into multiple candidate image sequences according to a preset sequence length and a preset step length, where the length of the candidate image sequence is equal to the length of the preset sequence, and adjacent candidate image sequences overlap
  • the number of images is equal to the difference between the preset sequence length and the preset step size.
  • the fifth determining module includes:
  • An eighth determining submodule configured to determine a matching score of each query image sequence of the query video and each candidate image sequence of the candidate video
  • a third calculation submodule configured to calculate an average value of the highest N matching scores among the matching scores of each query image sequence of the query video and each candidate image sequence of the candidate video to obtain the query video and the Matching scores of candidate videos, where N is a positive integer;
  • a ninth determining submodule is configured to determine a matching result between the query video and the candidate video based on a matching score of the query video and the candidate video.
  • an electronic device including:
  • Memory for storing processor-executable instructions
  • the processor is configured to execute the target matching method.
  • a computer-readable storage medium having computer program instructions stored thereon, the computer program instructions realizing the above-mentioned target matching method when executed by a processor.
  • the query image sequence is determined based on the self-expression feature vector of the query image sequence, the co-expression feature vector of the query image sequence, the self-expression feature vector of the candidate image sequence, and the co-expression feature vector of the candidate image sequence.
  • the similarity feature vector of the candidate image sequence and based on the similarity feature vector, determine the matching result between the query image sequence and the candidate image sequence, thereby improving the accuracy of target matching.
  • FIG. 1 illustrates a flowchart of a target matching method according to an embodiment of the present disclosure.
  • FIG. 2 shows an exemplary flowchart of step S12 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 3 illustrates an exemplary flowchart of step S121 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 4 illustrates an exemplary flowchart of step S122 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 5 illustrates an exemplary flowchart of step S1213 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 6 illustrates an exemplary flowchart of step S1223 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 7 illustrates an exemplary flowchart of step S13 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 8 illustrates an exemplary flowchart of step S131 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 9 illustrates an exemplary flowchart of step S132 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 10 illustrates an exemplary flowchart of step S1312 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 11 illustrates an exemplary flowchart of step S1322 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 12 illustrates an exemplary flowchart of step S14 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 13 illustrates an exemplary flowchart of step S15 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 14 illustrates an exemplary flowchart of a target matching method according to an embodiment of the present disclosure.
  • FIG. 15 illustrates an exemplary flowchart of step S28 of the target matching method according to an embodiment of the present disclosure.
  • FIG. 16 illustrates a block diagram of a target matching device according to an embodiment of the present disclosure.
  • FIG. 17 illustrates an exemplary block diagram of a target matching device according to an embodiment of the present disclosure.
  • Fig. 18 is a block diagram showing an electronic device 800 according to an exemplary embodiment.
  • Fig. 19 is a block diagram of an electronic device 1900 according to an exemplary embodiment.
  • exemplary means “serving as an example, embodiment, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as superior or better than other embodiments.
  • FIG. 1 illustrates a flowchart of a target matching method according to an embodiment of the present disclosure.
  • the embodiments of the present disclosure can be applied to fields such as intelligent video analysis or security monitoring.
  • the embodiments of the present disclosure can be combined with technologies such as pedestrian detection and pedestrian tracking, and can be applied to security monitoring systems at airports, stations, campuses, or shopping malls.
  • the method includes steps S11 to S15.
  • step S11 the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence are separately extracted, where the query image sequence includes the target to be matched.
  • the query image sequence may refer to an image sequence that requires target matching.
  • the candidate image sequence may refer to an image sequence in a database.
  • the database may contain multiple candidate image sequences, for example, the database may contain a large number of candidate image sequences.
  • the query image sequence may include only one target to be matched, or may include multiple targets to be matched.
  • the image sequence in the embodiment of the present disclosure may be a video, a video fragment, or another image sequence.
  • the number of frames of the query image sequence and the candidate image sequence may be different or the same.
  • the query image sequence contains T frames (ie, the query image sequence contains T images)
  • the candidate image sequence contains R frames (ie, the candidate image sequence contains R images), where both T and R are positive integers.
  • the feature vector of each frame in the query image sequence is extracted to obtain Among them, x t represents the feature vector of the t-th frame in the query image sequence, 1 ⁇ t ⁇ T; the feature vector of each frame in the candidate image sequence is extracted to obtain Among them, y r represents the feature vector of the r-th frame in the candidate image sequence, and 1 ⁇ r ⁇ R.
  • extracting the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence separately include: extracting the Feature vector and feature vector of each frame in the candidate image sequence.
  • the first child neural network may be a CNN (Convolutional Neural Network, Convolutional Neural Network).
  • a convolutional neural network with the same parameters can be used to extract the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence.
  • the method further includes: a first full connection through a first sub-neural network
  • the layer performs dimension reduction processing on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence to obtain the first dimensionality reduction feature vector of each frame in the query image sequence and each of the candidate image sequences The first dimensionality reduction feature vector of the frame.
  • the first dimensionality reduction feature vector of each frame in the query image sequence can be expressed as among them, Represents the first dimensionality reduction feature vector of the t-th frame in the query image sequence; the first dimensionality reduction feature vector of each frame in the candidate image sequence can be expressed as among them, Represents the first dimensionality reduction feature vector of the r-th frame in the candidate image sequence.
  • the dimension of the feature vector of each frame in the query image sequence is 2048
  • the dimension of the first dimension-reduced feature vector of each frame in the query image sequence is 128.
  • the number of dimensions is 2048
  • the dimension of the first dimensionality reduction feature vector of each frame in the candidate image sequence is 128.
  • the first fully connected layer may be denoted as fc-0.
  • step S12 the self-expression feature vector of the query image sequence and the self-expression feature vector of the candidate image sequence are determined based on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence.
  • the self-expression feature vector of the query image sequence may be determined based on the feature vector of each frame in the query image sequence; the self-expression feature of the candidate image sequence may be determined based on the feature vector of each frame in the candidate image sequence. vector.
  • the self-expression feature vector of the query image sequence may represent a feature vector determined only by the expression of the query image sequence, that is, the self-expression feature vector of the query image sequence is determined only by the expression of the query image sequence, and the candidate The expression of the image sequence is irrelevant;
  • the self-expression feature vector of the candidate image sequence can represent a feature vector determined only by the expression of the candidate image sequence, that is, the self-expression feature vector of the candidate image sequence is determined only by the expression of the candidate image sequence, and is related to the query image The expression of the sequence is irrelevant.
  • step S13 based on the feature vector of each frame in the query image sequence and the self-expression feature vector of the candidate image sequence, a collaborative expression feature vector of the query image sequence is determined, and the feature vector and query of each frame in the candidate image sequence are determined.
  • the self-expression feature vector of the image sequence determines the cooperative expression feature vector of the candidate image sequence.
  • the cooperative expression feature vector of the query image sequence may represent a feature vector determined by the expression of the query image sequence and the expression of the candidate image sequence, that is, the collaborative expression feature vector of the query image sequence is not only related to the query image sequence
  • the expression correlation of the candidate image sequence is also related to the expression of the candidate image sequence
  • the collaborative expression feature vector of the candidate image sequence can represent a feature vector determined by the expression of the candidate image sequence and the expression of the query image sequence, that is, the collaborative expression feature of the candidate image sequence
  • Vectors are not only related to the expression of candidate image sequences, but also to the expression of query image sequences.
  • step S14 the query image sequence and the candidate image sequence are determined based on the self-expression feature vector of the query image sequence, the collaborative expression feature vector of the query image sequence, the self-expression feature vector of the candidate image sequence, and the collaborative expression feature vector of the candidate image sequence Similarity feature vector.
  • the similarity feature vector of the query image sequence and the candidate image sequence may be used to determine the degree of similarity between the query image sequence and the candidate image sequence, so as to determine whether the query image sequence matches the candidate image sequence.
  • step S15 a matching result of the query image sequence and the candidate image sequence is determined based on the similarity feature vector.
  • the two matching image sequences may be image sequences of the same person captured from different shooting perspectives, or may be image sequences of the same person captured from the same shooting perspective.
  • the embodiments of the present disclosure determine a query image sequence and a candidate image sequence based on a self-expression feature vector of a query image sequence, a cooperative expression feature vector of a query image sequence, a self-expression feature vector of a candidate image sequence, and a cooperative expression feature vector of a candidate image sequence. Based on the similarity feature vector, and based on the similarity feature vector, determine the matching result between the query image sequence and the candidate image sequence, thereby improving the accuracy of target matching.
  • FIG. 2 shows an exemplary flowchart of step S12 of the target matching method according to an embodiment of the present disclosure.
  • step S12 may include steps S121 and S122.
  • step S121 the feature vector of each frame in the query image sequence and the first dimensionality reduction feature vector of each frame in the query image sequence are input to the second sub-neural network to determine the self-expressing feature vector of the query image sequence.
  • the second child neural network may be a SAN (Self Attention Subnetwork, an auto-expressor neural network based on the attention mechanism).
  • SAN Self Attention Subnetwork, an auto-expressor neural network based on the attention mechanism.
  • step S122 the feature vector of each frame in the candidate image sequence and the first dimensionality reduction feature vector of each frame in the candidate image sequence are input into the second sub-neural network to determine the self-expressing feature vector of the candidate image sequence.
  • the feature vector of each frame in the candidate image sequence can be And the first dimensionality reduction feature vector of each frame in the candidate image sequence Enter the second sub-neural network to determine the self-expression feature vector of the candidate image sequence
  • FIG. 3 illustrates an exemplary flowchart of step S121 of the target matching method according to an embodiment of the present disclosure.
  • step S121 may include steps S1211 and S1213.
  • step S1211 the feature vector of each frame in the query image sequence is reduced by the second fully connected layer of the second sub-neural network to obtain the second dimension-reduced feature vector of each frame in the query image sequence.
  • the second dimensionality reduction feature vector of each frame in the query image sequence can be expressed as among them, The second dimensionality reduction feature vector representing the t-th frame in the query image sequence.
  • the second fully connected layer can be denoted as fc-1.
  • the dimension of the second dimensionality reduction feature vector of each frame in the query image sequence is 128 dimensions.
  • step S1212 the second dimensionality reduction feature vector of each frame in the query image sequence is subjected to an average pooling process in the time dimension to obtain the overall feature vector of the query image sequence.
  • the overall feature vector of a query image sequence can be expressed as
  • step S1213 based on the second dimensionality reduction feature vector of each frame in the query image sequence, the overall feature vector of the query image sequence, and the first dimensionality reduction feature vector of each frame in the query image sequence, the self-determination of the query image sequence is determined.
  • the self-determination of the query image sequence is determined.
  • FIG. 4 illustrates an exemplary flowchart of step S122 of the target matching method according to an embodiment of the present disclosure.
  • step S122 may include steps S1221 and S1223.
  • step S1221 the feature vector of each frame in the candidate image sequence is reduced by the second fully-connected layer of the second sub-neural network to obtain the second reduced-dimensional feature vector of each frame in the candidate image sequence.
  • the dimension of the second dimensionality reduction feature vector of each frame in the candidate image sequence is 128 dimensions.
  • step S1222 the second dimensionality reduction feature vector of each frame in the candidate image sequence is subjected to an average pooling process in the time dimension to obtain the overall feature vector of the candidate image sequence.
  • the overall feature vector of a candidate image sequence can be expressed as
  • step S1223 based on the second dimensionality reduction feature vector of each frame in the candidate image sequence, the overall feature vector of the candidate image sequence, and the first dimensionality reduction feature vector of each frame in the candidate image sequence, the self- Express feature vectors.
  • FIG. 5 illustrates an exemplary flowchart of step S1213 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 5, step S1213 may include steps S12131 and S12132.
  • step S12131 the correlation between the second dimensionality-reduced feature vector of each frame in the query image sequence and the overall feature vector of the query image sequence is calculated through a parameterless correlation function to obtain the first correlation weight of each frame in the query image sequence.
  • the correlation between the second dimensionality-reduced feature vector of each frame in the query image sequence and the overall feature vector of the query image sequence can be calculated by the parameterless correlation function f () to obtain the first correlation of each frame in the query image sequence.
  • the parameterless correlation function f () can be calculated by dot multiplication. versus Relevance.
  • the embodiment of the present disclosure is based on a self-expression mechanism, and assigns a relevant weight to each frame of the query image sequence through the query image sequence's own expression.
  • step S12132 based on the first correlation weight of each frame in the query image sequence, the first dimensionality reduction feature vector of each frame in the query image sequence is weighted to obtain the self-expression feature vector of the query image sequence.
  • the self-expressing feature vector of a query image sequence can be expressed as among them,
  • the second dimensionality reduction feature vector of the t-th frame in the query image sequence Represents the overall feature vector of the query image sequence, Represents the first dimensionality reduction feature vector of the t-th frame in the query image sequence.
  • the first correlation weight includes a first normalized correlation weight
  • the first normalized correlation weight is obtained by performing a normalization process on the first correlation weight.
  • weighting the first dimensionality reduction feature vector of each frame in the query image sequence to obtain a self-expressing feature vector of the query image sequence including: Normalize the first correlation weight of each frame in the query image sequence to obtain the first normalized correlation weight of each frame in the query image sequence; based on the first normalization of each frame in the query image sequence
  • the correlation weight is used to weight the first dimensionality reduction feature vector of each frame in the query image sequence to obtain the self-expression feature vector of the query image sequence.
  • softmax may be used to normalize the first correlation weight of each frame in the query image sequence to obtain the first normalized correlation weight of each frame in the query image sequence.
  • FIG. 6 illustrates an exemplary flowchart of step S1223 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 6, step S1223 may include steps S12231 and S12232.
  • step S12231 the correlation between the second dimensionality-reduced feature vector of each frame in the candidate image sequence and the overall feature vector of the candidate image sequence is calculated through a parameterless correlation function to obtain the first correlation weight of each frame in the candidate image sequence.
  • the correlation between the second dimensionality-reduced feature vector of each frame in the candidate image sequence and the overall feature vector of the candidate image sequence can be calculated by the parameterless correlation function f () to obtain the first correlation of each frame in the candidate image sequence.
  • the parameterless correlation function f () can be calculated by dot multiplication. versus Relevance.
  • the embodiment of the present disclosure is based on a self-expression mechanism, and assigns a relevant weight to each frame of the candidate image sequence through its own expression of the candidate image sequence.
  • step S12232 based on the first correlation weight of each frame in the candidate image sequence, the first dimensionality reduction feature vector of each frame in the candidate image sequence is weighted to obtain the self-expression feature vector of the candidate image sequence.
  • the self-expressing feature vector of a candidate image sequence can be expressed as among them, The second dimensionality reduction feature vector of the r-th frame in the candidate image sequence, The overall feature vector representing the candidate image sequence, Represents the first dimensionality reduction feature vector of the r-th frame in the candidate image sequence.
  • the first correlation weight includes a first normalized correlation weight
  • the first normalized correlation weight is obtained by performing a normalization process on the first correlation weight.
  • weighting the first dimensionality reduction feature vector of each frame in the candidate image sequence to obtain a self-expression feature vector of the candidate image sequence including: Normalize the first correlation weight of each frame in the candidate image sequence to obtain the first normalized correlation weight of each frame in the candidate image sequence; based on the first normalization of each frame in the candidate image sequence
  • the correlation weight is used to weight the first dimensionality reduction feature vector of each frame in the candidate image sequence to obtain the self-expression feature vector of the candidate image sequence.
  • softmax may be used to normalize the first correlation weight of each frame in the candidate image sequence to obtain the first normalized correlation weight of each frame in the candidate image sequence.
  • FIG. 7 illustrates an exemplary flowchart of step S13 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 7, step S13 may include steps S131 and S132.
  • step S131 the feature vector of each frame in the query image sequence, the first dimensionality reduction feature vector of each frame in the query image sequence, and the self-expressing feature vector of the candidate image sequence are input into the third sub-neural network to obtain a query.
  • Cooperative expression feature vector of image sequence is input into the third sub-neural network to obtain a query.
  • the third sub-neural network may be CAN (Collaborative Attention Subnetwork, a cooperative expression sub-neural network based on attention mechanism).
  • step S132 the feature vector of each frame in the candidate image sequence, the first dimensionality reduction feature vector of each frame in the candidate image sequence, and the self-expressing feature vector of the query image sequence are input into a third sub-neural network to obtain a candidate.
  • Cooperative expression feature vector of image sequence is input into a third sub-neural network to obtain a candidate.
  • the feature vector of each frame in the candidate image sequence can be The first dimensionality reduction feature vector of each frame in the candidate image sequence
  • query self-expressing feature vectors of image sequences Enter the third sub-neural network to obtain the collaborative expression feature vector of the candidate image sequence
  • FIG. 8 illustrates an exemplary flowchart of step S131 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 8, step S131 may include steps S1311 and S1312.
  • step S1311 the feature vector of each frame in the query image sequence is reduced by the third fully-connected layer of the third sub-neural network to obtain the third dimension-reduced feature vector of each frame in the query image sequence.
  • the third dimensionality reduction feature vector of each frame in the query image sequence can be expressed as among them, Represents the third dimensionality reduction feature vector of the t-th frame in the query image sequence.
  • the dimension of the third dimensionality reduction feature vector of each frame in the query image sequence is 128 dimensions.
  • the third fully connected layer can be represented as fc-2.
  • step S1312 the query image sequence is obtained based on the third dimension reduction feature vector of each frame in the query image sequence, the self-expression feature vector of the candidate image sequence, and the first dimension reduction feature vector of each frame in the query image sequence. Co-express feature vectors.
  • FIG. 9 illustrates an exemplary flowchart of step S132 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 9, step S132 may include steps S1321 and S1322.
  • step S1321 the feature vector of each frame in the candidate image sequence is reduced by the third fully-connected layer of the third sub-neural network to obtain the third reduced-dimensional feature vector of each frame in the candidate image sequence.
  • the third dimensionality reduction feature vector of each frame in the candidate image sequence can be expressed as among them, Represents the third dimensionality reduction feature vector of the r-th frame in the candidate image sequence.
  • the dimension of the third dimensionality reduction feature vector of each frame in the candidate image sequence is 128 dimensions.
  • step S1322 the candidate image sequence is obtained based on the third dimensionality reduction feature vector of each frame in the candidate image sequence, the self-expression feature vector of the query image sequence, and the first dimensionality reduction feature vector of each frame in the candidate image sequence. Co-express feature vectors.
  • FIG. 10 illustrates an exemplary flowchart of step S1312 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 10, step S1312 may include steps S13121 and S13122.
  • step S13121 the correlation between the third dimensionality-reduced feature vector of each frame in the query image sequence and the self-expressing feature vector of the candidate image sequence is calculated through a parameterless correlation function to obtain the second correlation of each frame in the query image sequence. Weights.
  • the second correlation weight of the t-th frame in the query image sequence can be expressed as
  • the embodiment of the present disclosure is based on a cooperative expression mechanism, and assigns a relevant weight to each frame of the query image sequence through the expression of the candidate image sequence and the query image sequence's own expression.
  • step S13122 based on the second correlation weight of each frame in the query image sequence, the first dimensionality reduction feature vector of each frame in the query image sequence is weighted to obtain the collaborative expression feature vector of the query image sequence.
  • the collaborative expression feature vector of a query image sequence can be expressed as
  • the second related weight includes a second normalized related weight
  • the second normalized related weight is obtained by normalizing the second related weight.
  • weighting the first dimensionality reduction feature vector of each frame in the query image sequence to obtain a collaborative expression feature vector of the query image sequence including: Normalize the second correlation weight of each frame in the query image sequence to obtain the second normalized correlation weight of each frame in the query image sequence; based on the second normalization of each frame in the query image sequence
  • the correlation weight is used to weight the first dimensionality reduction feature vector of each frame in the query image sequence to obtain the collaborative expression feature vector of the query image sequence.
  • FIG. 11 illustrates an exemplary flowchart of step S1322 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 11, step S1322 may include steps S13221 and S13222.
  • step S13221 the correlation between the third dimensionality-reduced feature vector of each frame in the candidate image sequence and the self-expression feature vector of the query image sequence is calculated by using a parameterless correlation function to obtain the second correlation of each frame in the candidate image sequence. Weights.
  • the second correlation weight of the r-th frame in the candidate image sequence can be expressed as
  • the embodiment of the present disclosure is based on a cooperative expression mechanism, and assigns a relevant weight to each frame of the candidate image sequence by querying the expression of the image sequence and the own expression of the candidate image sequence.
  • step S13222 based on the second correlation weight of each frame in the candidate image sequence, the first dimensionality reduction feature vector of each frame in the candidate image sequence is weighted to obtain the collaborative expression feature vector of the candidate image sequence.
  • the collaborative expression feature vector of a candidate image sequence can be expressed as
  • the second related weight includes a second normalized related weight
  • the second normalized related weight is obtained by normalizing the second related weight.
  • weighting the first dimensionality reduction feature vector of each frame in the candidate image sequence to obtain a collaborative expression feature vector of the candidate image sequence including: Normalize the second correlation weight of each frame in the candidate image sequence to obtain the second normalized correlation weight of each frame in the candidate image sequence; based on the second normalization of each frame in the candidate image sequence
  • the correlation weight is used to weight the first dimensionality reduction feature vector of each frame in the candidate image sequence to obtain the collaborative expression feature vector of the candidate image sequence.
  • the second sub-neural network and the third sub-neural network are based on the self-expression mechanism and the cooperative expression mechanism, and query each frame of the image sequence and the candidate image by querying the expression of the image sequence and the expression of the candidate image sequence.
  • Each frame of the sequence is assigned a relevant weight.
  • the second child neural network and the third child neural network use this non-parametric self-expression and cooperative expression to implicitly align the query image sequence and the candidate image sequence to select a more discriminative frame for two images. Sequence for expression. Since the second sub-neural network and the third sub-neural network are non-parametric, the query image sequence and the candidate image sequence are allowed to have different lengths. Therefore, the target matching method provided by the embodiment of the present disclosure has high flexibility and can be widely applied.
  • FIG. 12 illustrates an exemplary flowchart of step S14 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 12, step S14 may include steps S141 and S143.
  • step S141 the difference between the self-expression feature vector of the query image sequence and the co-expression feature vector of the candidate image sequence is calculated to obtain a first difference vector.
  • the first difference vector is
  • step S142 the difference between the self-expression feature vector of the candidate image sequence and the co-expression feature vector of the query image sequence is calculated to obtain a second difference vector.
  • the second difference vector is
  • step S143 based on the first difference vector and the second difference vector, a similarity feature vector of the query image sequence and the candidate image sequence is obtained.
  • obtaining the similarity feature vector of the query image sequence and the candidate image sequence based on the first difference vector and the second difference vector includes: calculating a sum of the first difference vector and the second difference vector to obtain Query the similarity feature vector between the image sequence and the candidate image sequence. For example, query similarity feature vectors of image sequences and candidate image sequences
  • obtaining a similarity feature vector between the query image sequence and the candidate image sequence based on the first difference vector and the second difference vector including: calculating corresponding bits of the first difference vector and the second difference vector.
  • the product of the elements of is used to obtain the similarity feature vector of the query image sequence and the candidate image sequence.
  • FIG. 13 illustrates an exemplary flowchart of step S15 of the target matching method according to an embodiment of the present disclosure. As shown in FIG. 13, step S15 may include steps S151 and S152.
  • step S151 the similarity feature vector of the query image sequence and the candidate image sequence is input to the fourth fully connected layer to obtain a matching score between the query image sequence and the candidate image sequence.
  • the fourth fully connected layer can be represented as fc-3.
  • step S152 based on the matching score of the query image sequence and the candidate image sequence, a matching result of the query image sequence and the candidate image sequence is determined.
  • the matching score between the query image sequence and the candidate image sequence is greater than the score threshold, it can be determined that the matching result between the query image sequence and the candidate image sequence is that the query image sequence matches the candidate image sequence; if the query image sequence matches the candidate image sequence, If the matching score is less than or equal to the score threshold, it can be determined that the matching result between the query image sequence and the candidate image sequence is that the query image sequence does not match the candidate image sequence.
  • the method further includes: using the same pair of labeled data and binary cross entropy based on the matching score of the query image sequence and the candidate image sequence. Loss function to optimize network parameters.
  • N the number of query image sequence and candidate image sequence pairs in the training set
  • the training image sequence can be segmented to generate rich query image sequence and candidate image sequence pairs, thereby effectively improving optimization efficiency and improving the robustness of the network model. Thereby, matching accuracy can be improved.
  • FIG. 14 illustrates an exemplary flowchart of a target matching method according to an embodiment of the present disclosure. As shown in FIG. 14, the method may include steps S21 to S28.
  • step S21 the query video is divided into a plurality of query image sequences.
  • segmenting the query video into multiple query image sequences includes: segmenting the query video into multiple query image sequences according to a preset sequence length and a preset step size, where the query images The length of the sequence is equal to the preset sequence length, and the number of overlapping images between adjacent query image sequences is equal to the difference between the preset sequence length and the preset step size.
  • step S22 the candidate video is segmented into a plurality of candidate image sequences.
  • segmenting the candidate video into multiple candidate image sequences includes: segmenting the candidate video into multiple candidate image sequences according to a preset sequence length and a preset step length, where the candidate images The length of the sequence is equal to the preset sequence length, and the number of overlapping images between adjacent candidate image sequences is equal to the difference between the preset sequence length and the preset step size.
  • step S23 the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence are separately extracted, where the query image sequence includes the target to be matched.
  • step S23 refer to the description of step S11 above.
  • step S24 the self-expression feature vector of the query image sequence and the self-expression feature vector of the candidate image sequence are determined based on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence.
  • step S24 refer to the description of step S12 above.
  • step S25 based on the feature vector of each frame in the query image sequence and the self-expression feature vector of the candidate image sequence, a collaborative expression feature vector of the query image sequence is determined, and the feature vector and query based on each frame in the candidate image sequence The self-expression feature vector of the image sequence determines the cooperative expression feature vector of the candidate image sequence.
  • step S25 refer to the description of step S13 above.
  • step S26 the query image sequence and the candidate image sequence are determined based on the self-expression feature vector of the query image sequence, the co-expression feature vector of the query image sequence, the self-expression feature vector of the candidate image sequence, and the co-expression feature vector of the candidate image sequence. Similarity feature vector.
  • step S26 refer to the description of step S14 above.
  • step S27 a matching result of the query image sequence and the candidate image sequence is determined based on the similarity feature vector.
  • step S27 refer to the description of step S15 above.
  • step S28 based on the matching result between the query image sequence of the query video and the candidate image sequence of the candidate video, a matching result of the query video and the candidate video is determined.
  • FIG. 15 illustrates an exemplary flowchart of step S28 of the target matching method according to an embodiment of the present disclosure.
  • step S28 may include steps S281 to S283.
  • step S281 a matching score of each query image sequence of the query video and each candidate image sequence of the candidate video is determined.
  • step S282 the average value of the highest N matching scores among the matching scores of each query image sequence of the query video and each candidate image sequence of the candidate video is calculated to obtain the matching score of the query video and the candidate video, where N is positive Integer.
  • step S283 a matching result of the query video and the candidate video is determined based on the matching score of the query video and the candidate video.
  • the matching score between the query video and the candidate video is greater than the score threshold, it can be determined that the matching result between the query video and the candidate video is that the query video matches the candidate video; if the query video matches the candidate video, If the score is less than or equal to the score threshold, it can be determined that the matching result between the query video and the candidate video is that the query video does not match the candidate video.
  • the target matching method provided by the embodiment of the present disclosure can filter out more discriminating key frames in the image sequence, and use multiple key frames to express the image sequence, thereby improving the discrimination ability; the embodiment of the present disclosure proposes more effective
  • the time-domain modeling method captures the dynamic change information of consecutive frames and improves the expression ability of the model.
  • the embodiment of the present disclosure proposes a more effective distance measurement method, which reduces the distance between feature expressions of the same person, increasing Increase the distance between the character expressions of different characters.
  • the target matching method provided by the embodiment of the present disclosure can still obtain more accurate target matching results under the conditions of poor lighting conditions, severe occlusion, poor viewing angle, or severe background interference.
  • the embodiments of the present disclosure can help improve the effects of human detection and / or pedestrian tracking. Utilizing the embodiments of the present disclosure, it is possible to perform better cross-camera search and tracking on specific pedestrians (such as criminal suspects, missing children, etc.) in intelligent video surveillance.
  • the present disclosure also provides a target matching device, an electronic device, a computer-readable storage medium, and a program, all of which can be used to implement any one of the target matching methods provided by the present disclosure, the corresponding technical solutions and descriptions, and the corresponding records in the method section ,No longer.
  • FIG. 16 illustrates a block diagram of a target matching device according to an embodiment of the present disclosure.
  • the device includes: an extraction module 31 for extracting a feature vector of each frame in a query image sequence and a feature vector of each frame in a candidate image sequence, wherein the query image sequence includes a target to be matched;
  • the first determining module 32 is configured to determine the self-expression feature vector of the query image sequence and the self-expression feature vector of the candidate image sequence based on the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence, respectively.
  • a second determining module 33 configured to determine a cooperative expression feature vector of the query image sequence based on the feature vector of each frame in the query image sequence and a self-expression feature vector of the candidate image sequence, and based on the The feature vector and the self-expressing feature vector of the query image sequence determine the cooperative expression feature vector of the candidate image sequence;
  • the third determination module 34 is used to base the self-expression feature vector of the query image sequence, the collaborative expression feature vector of the query image sequence, the candidate Self-expressing feature vectors of image sequences and collaborative expression of candidate image sequences
  • the feature vector determines a similarity feature vector between the query image sequence and the candidate image sequence;
  • a fourth determination module 35 is configured to determine a matching result between the query image sequence and the candidate image sequence based on the similarity feature vector.
  • the extraction module 31 is configured to extract the feature vector of each frame in the query image sequence and the feature vector of each frame in the candidate image sequence through the first sub-neural network.
  • FIG. 17 illustrates an exemplary block diagram of a target matching device according to an embodiment of the present disclosure. As shown in Figure 17:
  • the apparatus further includes: a dimensionality reduction module 36, configured to query the feature vector of each frame in the image sequence and each of the candidate image sequences through the first fully connected layer of the first sub-neural network.
  • the feature vector of one frame is subjected to dimension reduction processing to obtain the first dimension-reduced feature vector of each frame in the query image sequence and the first dimension-reduced feature vector of each frame in the candidate image sequence.
  • the first determining module 32 includes: a first determining submodule 321, configured to convert a feature vector of each frame in the query image sequence and a first dimension reduction feature of each frame in the query image sequence The vector is input into the second sub-neural network to determine the self-expressing feature vector of the query image sequence; the second determination sub-module 322 is configured to combine the feature vector of each frame in the candidate image sequence and the first of each frame in the candidate image sequence The dimensionality-reduced feature vector is input into the second sub-neural network to determine the self-expressing feature vector of the candidate image sequence.
  • the first determining sub-module 321 includes: a first dimension reduction unit, configured to reduce the feature vector of each frame in the query image sequence through the second fully connected layer of the second sub-neural network Dimensional processing to obtain the second dimensionality reduction feature vector of each frame in the query image sequence; a first average pooling unit for averaging the second dimension reduction feature vector of each frame in the query image sequence over the time dimension Processing to obtain the overall feature vector of the query image sequence; a first determining unit, configured to: based on the second dimensionality reduction feature vector of each frame in the query image sequence, the overall feature vector of the query image sequence, and the The first dimensionality reduction feature vector determines the self-expression feature vector of the query image sequence.
  • the second determining submodule 322 includes: a second dimension reduction unit, configured to reduce the feature vector of each frame in the candidate image sequence through the second fully connected layer of the second sub-neural network. Dimension processing to obtain a second dimensionality reduction feature vector of each frame in the candidate image sequence; a second average pooling unit for averaging the second dimensionality reduction feature vector of each frame in the candidate image sequence over the time dimension Processing to obtain the overall feature vector of the candidate image sequence; a second determining unit, configured to be based on the second dimensionality reduction feature vector of each frame in the candidate image sequence, the overall feature vector of the candidate image sequence, and the The first dimensionality reduction feature vector determines the self-expression feature vector of the candidate image sequence.
  • the first determining unit includes: a first calculation subunit, configured to calculate a second dimension-reduced feature vector of each frame in the query image sequence and the overall feature of the query image sequence through a parameterless correlation function.
  • the first dimensionality reduction feature vector is weighted to obtain a self-expressing feature vector of the query image sequence.
  • the second determination unit includes: a second calculation subunit, configured to calculate a second dimension-reduced feature vector of each frame in the candidate image sequence and an overall feature of the candidate image sequence by using a parameterless correlation function.
  • the first dimensionality reduction feature vector is weighted to obtain a self-expression feature vector of the candidate image sequence.
  • the first correlation weight includes a first normalized correlation weight
  • the first normalized correlation weight is obtained by performing a normalization process on the first correlation weight.
  • the second determination module 33 includes: a third determination submodule 331, configured to query a feature vector of each frame in the query image sequence, and query the first dimension reduction feature of each frame in the image sequence
  • the vector and the self-expressing feature vector of the candidate image sequence are input into the third sub-neural network to obtain the collaborative expression feature vector of the query image sequence
  • the fourth determining submodule 332 is configured to use the feature vector and candidate of each frame in the candidate image sequence
  • the first dimensionality-reduced feature vector of each frame in the image sequence and the self-expression feature vector of the query image sequence are input into the third sub-neural network to obtain the collaborative expression feature vector of the candidate image sequence.
  • the third determining sub-module 331 includes a third dimension reduction unit, configured to reduce the feature vector of each frame in the query image sequence through the third fully connected layer of the third sub-neural network. Dimensional processing to obtain the third dimensionality reduction feature vector of each frame in the query image sequence; a third determination unit, which is based on the third dimensionality reduction feature vector of each frame in the query image sequence and the self-expression feature vector of the candidate image sequence And querying the first dimensionality reduction feature vector of each frame in the image sequence to obtain the collaborative expression feature vector of the query image sequence; the fourth determining submodule 332 includes a fourth dimensionality reduction unit for The three fully connected layers perform a dimensionality reduction process on the feature vectors of each frame in the candidate image sequence to obtain a third dimensionality reduction feature vector of each frame in the candidate image sequence; a fourth determination unit is configured to be based on each of the candidate image sequences The third dimensionality-reduced feature vector of the frame, the self-expression feature vector of the query
  • the third determination unit includes: a third calculation subunit, configured to calculate, through a parameterless correlation function, the third dimension-reduced feature vector of each frame in the query image sequence and the self-expression of the candidate image sequence The correlation degree of the feature vector to obtain the second correlation weight of each frame in the query image sequence; the third weighting subunit is used for each frame in the query image sequence based on the second correlation weight of each frame in the query image sequence The weighted first dimensionality reduction feature vector is weighted to obtain the collaborative expression feature vector of the query image sequence.
  • the fourth determination unit includes: a fourth calculation subunit, configured to calculate, through a parameterless correlation function, a third dimension-reduced feature vector of each frame in the candidate image sequence and a self-expression of the query image sequence The correlation degree of the feature vector to obtain the second correlation weight of each frame in the candidate image sequence; the fourth weighting subunit is used for each frame in the candidate image sequence based on the second correlation weight of each frame in the candidate image sequence The weighted first dimensionality reduction feature vector is weighted to obtain a collaborative expression feature vector of the candidate image sequence.
  • the second related weight includes a second normalized related weight
  • the second normalized related weight is obtained by normalizing the second related weight
  • the third determination module 34 includes: a first calculation submodule 341, configured to calculate a difference between a self-expression feature vector of a query image sequence and a co-expression feature vector of a candidate image sequence to obtain a first difference
  • a fifth determination sub-module 343 based on the first The difference vector and the second difference vector are used to obtain the similarity feature vector of the query image sequence and the candidate image sequence.
  • the fifth determination sub-module 343 includes: a first calculation unit, configured to calculate a sum of the first difference vector and the second difference vector, to obtain a similarity feature vector of the query image sequence and the candidate image sequence Or, a second calculation unit, configured to calculate a product of elements of corresponding bits of the first difference vector and the second difference vector, to obtain a similarity feature vector of the query image sequence and the candidate image sequence.
  • the fourth determination module 35 includes: a sixth determination submodule 351, configured to input the similarity feature vector of the query image sequence and the candidate image sequence into the fourth fully connected layer to obtain the query image sequence and A matching score of the candidate image sequence; a seventh determination submodule 352 is configured to determine a matching result of the query image sequence and the candidate image sequence based on the matching score of the query image sequence and the candidate image sequence.
  • the device further includes: an optimization module 37 for optimizing network parameters based on the matching score of the query image sequence and the candidate image sequence, using the same pair of labeled data and a binary cross-entropy loss function.
  • the device further includes a first segmentation module 38 for segmenting the query video into multiple query image sequences, and a second segmentation module 39 for segmenting the candidate video into A plurality of candidate image sequences; a fifth determination module 30 is configured to determine a matching result between the query video and the candidate video based on a matching result between the query image sequence of the query video and the candidate image sequence of the candidate video.
  • the first segmentation module 38 is configured to segment the query video into multiple query image sequences according to a preset sequence length and a preset step length, where the length of the query image sequence is equal to the Set the sequence length, the number of overlapping images between adjacent query image sequences is equal to the difference between the preset sequence length and the preset step size;
  • the second segmentation module 39 is used to: according to the preset sequence length and the preset step size, The candidate video is divided into multiple candidate image sequences, where the length of the candidate image sequence is equal to the preset sequence length, and the number of overlapping images between adjacent candidate image sequences is equal to the difference between the preset sequence length and the preset step size.
  • the fifth determination module 30 includes: an eighth determination submodule 301, configured to determine a matching score between each query image sequence of the query video and each candidate image sequence of the candidate video; and a third calculation submodule 302. Calculate the average of the highest N matching scores of the matching scores of each query image sequence of the query video and each candidate image sequence of the candidate video to obtain the matching score of the query video and the candidate video, where N is a positive integer.
  • a ninth determining sub-module 303 configured to determine a matching result between the query video and the candidate video based on the matching score of the query video and the candidate video.
  • the embodiments of the present disclosure determine a query image sequence and a candidate image sequence based on a self-expression feature vector of a query image sequence, a cooperative expression feature vector of a query image sequence, a self-expression feature vector of a candidate image sequence, and a cooperative expression feature vector of a candidate image sequence. Based on the similarity feature vector, and based on the similarity feature vector, determine the matching result between the query image sequence and the candidate image sequence, thereby improving the accuracy of target matching.
  • An embodiment of the present disclosure also provides a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions implement the above method when executed by a processor.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure further provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured as the foregoing method.
  • the electronic device may be provided as a terminal, a server, or other forms of devices.
  • Fig. 18 is a block diagram showing an electronic device 800 according to an exemplary embodiment.
  • the electronic device 800 may be a mobile phone, a computer, a digital broadcasting terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, and other terminals.
  • the electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power component 806, a multimedia component 808, an audio component 810, an input / output (I / O) interface 812, and a sensor component 814 , And communication component 816.
  • the processing component 802 generally controls overall operations of the electronic device 800, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the method described above.
  • the processing component 802 may include one or more modules to facilitate the interaction between the processing component 802 and other components.
  • the processing component 802 may include a multimedia module to facilitate the interaction between the multimedia component 808 and the processing component 802.
  • the memory 804 is configured to store various types of data to support operation at the electronic device 800. Examples of these data include instructions for any application or method for operating on the electronic device 800, contact data, phone book data, messages, pictures, videos, and the like.
  • the memory 804 may be implemented by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), Programming read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read-only memory
  • EPROM Programming read-only memory
  • PROM programmable read-only memory
  • ROM read-only memory
  • magnetic memory flash memory
  • flash memory magnetic disk or optical disk.
  • the power component 806 provides power to various components of the electronic device 800.
  • the power component 806 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power for the electronic device 800.
  • the multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and a user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive an input signal from a user.
  • the touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure related to the touch or slide operation.
  • the multimedia component 808 includes a front camera and / or a rear camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and / or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capabilities.
  • the audio component 810 is configured to output and / or input audio signals.
  • the audio component 810 includes a microphone (MIC).
  • the microphone is configured to receive an external audio signal.
  • the received audio signal may be further stored in the memory 804 or transmitted via the communication component 816.
  • the audio component 810 further includes a speaker for outputting audio signals.
  • the I / O interface 812 provides an interface between the processing component 802 and a peripheral interface module.
  • the peripheral interface module may be a keyboard, a click wheel, a button, or the like. These buttons can include, but are not limited to: a home button, a volume button, a start button, and a lock button.
  • the sensor component 814 includes one or more sensors for providing various aspects of the state evaluation of the electronic device 800.
  • the sensor component 814 can detect the on / off state of the electronic device 800, and the relative positioning of the components.
  • the component is the display and keypad of the electronic device 800.
  • the sensor component 814 can also detect the electronic device 800 or an electronic device 800.
  • the position of the component changes, the presence or absence of the user's contact with the electronic device 800, the orientation or acceleration / deceleration of the electronic device 800, and the temperature change of the electronic device 800.
  • the sensor component 814 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact.
  • the sensor component 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications.
  • the sensor component 814 may further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • the communication component 816 is configured to facilitate wired or wireless communication between the electronic device 800 and other devices.
  • the electronic device 800 can access a wireless network based on a communication standard, such as WiFi, 2G, or 3G, or a combination thereof.
  • the communication component 816 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 further includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra wideband
  • Bluetooth Bluetooth
  • the electronic device 800 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), Implementation of a programming gate array (FPGA), controller, microcontroller, microprocessor, or other electronic component to perform the above method.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA programming gate array
  • controller microcontroller, microprocessor, or other electronic component to perform the above method.
  • a non-volatile computer-readable storage medium such as a memory 804 including computer program instructions, and the computer program instructions may be executed by the processor 820 of the electronic device 800 to complete the foregoing method.
  • Fig. 19 is a block diagram of an electronic device 1900 according to an exemplary embodiment.
  • the electronic device 1900 may be provided as a server.
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932, for storing instructions executable by the processing component 1922, such as an application program.
  • the application program stored in the memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the method described above.
  • the electronic device 1900 may further include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output (I / O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as Windows ServerTM, Mac OSXTM, UnixTM, LinuxTM, FreeBSDTM, or the like.
  • a non-volatile computer-readable storage medium such as a memory 1932 including computer program instructions, and the computer program instructions may be executed by the processing component 1922 of the electronic device 1900 to complete the above method.
  • the present disclosure may be a system, method, and / or computer program product.
  • the computer program product may include a computer-readable storage medium having computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.
  • the computer-readable storage medium may be a tangible device that can hold and store instructions used by the instruction execution device.
  • the computer-readable storage medium may be, for example, but not limited to, an electric storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) Or flash memory), static random access memory (SRAM), portable compact disc read only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanical encoding device, such as a printer with instructions stored thereon A protruding structure in the hole card or groove, and any suitable combination of the above.
  • RAM random access memory
  • ROM read-only memory
  • EPROM erasable programmable read-only memory
  • flash memory flash memory
  • SRAM static random access memory
  • CD-ROM compact disc read only memory
  • DVD digital versatile disc
  • memory stick floppy disk
  • mechanical encoding device such as a printer with instructions stored thereon A protruding structure in the hole card or groove, and any suitable combination of the above.
  • Computer-readable storage media used herein are not to be interpreted as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (for example, light pulses through fiber optic cables), or via electrical wires Electrical signal transmitted.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers.
  • the network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device .
  • Computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or in one or more programming languages.
  • the programming languages include object-oriented programming languages—such as Smalltalk, C ++, and the like—and conventional procedural programming languages—such as "C” or similar programming languages.
  • Computer-readable program instructions may be executed entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer, partly on a remote computer, or entirely on a remote computer or server carried out.
  • the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as through an Internet service provider using the Internet connection).
  • electronic circuits such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) are personalized by using state information of computer-readable program instructions.
  • the electronic circuits may Computer-readable program instructions are executed to implement various aspects of the present disclosure.
  • These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing device, thereby producing a machine such that, when executed by a processor of a computer or other programmable data processing device , Means for implementing the functions / actions specified in one or more blocks in the flowcharts and / or block diagrams.
  • These computer-readable program instructions may also be stored in a computer-readable storage medium, and these instructions cause a computer, a programmable data processing apparatus, and / or other devices to work in a specific manner.
  • a computer-readable medium storing instructions includes: An article of manufacture that includes instructions to implement various aspects of the functions / acts specified in one or more blocks in the flowcharts and / or block diagrams.
  • Computer-readable program instructions can also be loaded onto a computer, other programmable data processing device, or other device, so that a series of operating steps can be performed on the computer, other programmable data processing device, or other device to produce a computer-implemented process , So that the instructions executed on the computer, other programmable data processing apparatus, or other equipment can implement the functions / actions specified in one or more blocks in the flowchart and / or block diagram.
  • each block in the flowchart or block diagram may represent a module, a program segment, or a part of an instruction that contains one or more components for implementing a specified logical function.
  • Executable instructions may also occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved.
  • each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts can be implemented in a dedicated hardware-based system that performs the specified function or action. , Or it can be implemented with a combination of dedicated hardware and computer instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Medical Informatics (AREA)
  • Library & Information Science (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种目标匹配方法及装置、电子设备和存储介质。该方法包括：提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量；基于查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，确定查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量和候选图像序列的协同表达特征向量；基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量和候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量；基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果。能够提高目标匹配的准确性。

Description

目标匹配方法及装置、电子设备和存储介质
本申请要求在2018年6月15日提交中国专利局、申请号为201810621959.5、申请名称为“目标匹配方法及装置、电子设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开涉及计算机视觉技术领域，尤其涉及一种目标匹配方法及装置、电子设备和存储介质。
背景技术
目标匹配是指返回数据库中与查询视频或查询图像具有相同目标的视频或图像。目标匹配技术广泛地应用于机场、车站、校园和商场等场所的安防监控系统中。相关技术中，目标匹配的准确性较低。
发明内容
本公开提出了一种目标匹配技术方案。
根据本公开的一方面，提供了一种目标匹配方法，包括：
分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，所述查询图像序列包含待匹配目标；
分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量；
基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量；
基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，确定所述查询图像序列与所述候选图像序列的相似性特征向量；
基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果。
在一种可能的实现方式中，分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，包括：
通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
在一种可能的实现方式中，在提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量之后，所述方法还包括：
通过第一子神经网络的第一全连接层对所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第一降维特征向量和所述候选图像序列中每一帧的第一降维特征向量。
在一种可能的实现方式中，分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量包括：
将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量；
将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量，包括：
通过所述第二子神经网络的第二全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第二降维特征向量；
将所述查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述查询图像序列的整体特征向量；
基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的整体特征向量以及所述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量。
在一种可能的实现方式中，将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，得到所述候选图像序列的自表达特征向量，包括：
通过所述第二子神经网络的第二全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第二降维特征向量；
将所述候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述候选图像序列的整体特征向量；
基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的整体特征向量以及所述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量，包括：
通过无参数相关函数计算所述查询图像序列中每一帧的第二降维特征向量与所述查询图像序列的整体特征向量的相关度，得到所述查询图像序列中每一帧的第一相关权重；
基于所述查询图像序列中每一帧的第一相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的自表达特征向量。
在一种可能的实现方式中，基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量，包括：
通过无参数相关函数计算所述候选图像序列中每一帧的第二降维特征向量与所述候选图像序列的整体特征向量的相关度，得到所述候选图像序列中每一帧的第一相关权重；
基于所述候选图像序列中每一帧的第一相关权重，对所述候选图像序列中每一帧的第一降维特征 向量进行加权，得到所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，所述第一相关权重包括第一归一化相关权重，所述第一归一化相关权重是对所述第一相关权重进行归一化处理得到的。
在一种可能的实现方式中，基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量，包括：
将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量；
将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量，包括：
通过所述第三子神经网络的第三全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第三降维特征向量；
基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量；
将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量，包括：
通过所述第三子神经网络的第三全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第三降维特征向量；
基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量，包括：
通过无参数相关函数计算所述查询图像序列中每一帧的第三降维特征向量与所述候选图像序列的自表达特征向量的相关度，得到所述查询图像序列中每一帧的第二相关权重；
基于所述查询图像序列中每一帧的第二相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的协同表达特征向量。
在一种可能的实现方式中，基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量，包括：
通过无参数相关函数计算所述候选图像序列中每一帧的第三降维特征向量与所述查询图像序列的自表达特征向量的相关度，得到所述候选图像序列中每一帧的第二相关权重；
基于所述候选图像序列中每一帧的第二相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，所述第二相关权重包括第二归一化相关权重，所述第二归一化相关权重是对所述第二相关权重进行归一化处理得到的。
在一种可能的实现方式中，基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量，包括：
计算所述查询图像序列的自表达特征向量与所述候选图像序列的协同表达特征向量之差，得到第一差向量；
计算所述候选图像序列的自表达特征向量与所述查询图像序列的协同表达特征向量之差，得到第二差向量；
基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
在一种可能的实现方式中，基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量，包括：
计算所述第一差向量与所述第二差向量之和，得到所述查询图像序列与所述候选图像序列的相似性特征向量；或者，
计算所述第一差向量与所述第二差向量的相应位的元素的乘积，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
在一种可能的实现方式中，基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果，包括：
将所述查询图像序列与所述候选图像序列的相似性特征向量输入第四全连接层，得到所述查询图像序列与所述候选图像序列的匹配分数；
基于所述查询图像序列与所述候选图像序列的匹配分数，确定所述查询图像序列与所述候选图像序列的匹配结果。
在一种可能的实现方式中，在得到所述查询图像序列与所述候选图像序列的匹配分数之后，所述方法还包括：
基于所述查询图像序列与所述候选图像序列的匹配分数，采用同对标注数据和二元交叉熵损失函数，优化网络参数。
在一种可能的实现方式中，在提取查询图像序列中每一帧的特征向量之前，所述方法还包括：
将查询视频切分为多个查询图像序列；
将候选视频切分为多个候选图像序列；
在确定所述查询图像序列与所述候选图像序列的匹配结果之后，所述方法还包括：
基于所述查询视频的查询图像序列与所述候选视频的候选图像序列的匹配结果，确定所述查询视 频与所述候选视频的匹配结果。
在一种可能的实现方式中，将查询视频切分为多个查询图像序列，包括：
按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，所述查询图像序列的长度等于所述预设序列长度，相邻的查询图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差；
将候选视频切分为多个候选图像序列，包括：
按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，所述候选图像序列的长度等于所述预设序列长度，相邻的候选图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差。
在一种可能的实现方式中，基于所述查询视频的查询图像序列与所述候选视频的候选图像序列的匹配结果，确定所述查询视频与所述候选视频的匹配结果，包括：
确定所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数；
计算所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到所述查询视频与所述候选视频的匹配分数，其中，N为正整数；
基于所述查询视频与所述候选视频的匹配分数，确定所述查询视频与所述候选视频的匹配结果。
根据本公开的一方面，提供了一种目标匹配装置，包括：
提取模块，用于分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，所述查询图像序列包含待匹配目标；
第一确定模块，用于分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量；
第二确定模块，用于基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量；
第三确定模块，用于基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，确定所述查询图像序列与所述候选图像序列的相似性特征向量；
第四确定模块，用于基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果。
在一种可能的实现方式中，所述提取模块用于：
通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
在一种可能的实现方式中，所述装置还包括：
降维模块，用于通过第一子神经网络的第一全连接层对所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第一降维特征向量和所述候选图像序列中每一帧的第一降维特征向量。
在一种可能的实现方式中，所述第一确定模块包括：
第一确定子模块，用于将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量；
第二确定子模块，用于将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，所述第一确定子模块包括：
第一降维单元，用于通过所述第二子神经网络的第二全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第二降维特征向量；
第一平均池化单元，用于将所述查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述查询图像序列的整体特征向量；
第一确定单元，用于基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的整体特征向量以及所述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量。
在一种可能的实现方式中，所述第二确定子模块包括：
第二降维单元，用于通过所述第二子神经网络的第二全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第二降维特征向量；
第二平均池化单元，用于将所述候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述候选图像序列的整体特征向量；
第二确定单元，用于基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，所述第一确定单元包括：
第一计算子单元，用于通过无参数相关函数计算所述查询图像序列中每一帧的第二降维特征向量与所述查询图像序列的整体特征向量的相关度，得到所述查询图像序列中每一帧的第一相关权重；
第一加权子单元，用于基于所述查询图像序列中每一帧的第一相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的自表达特征向量。
在一种可能的实现方式中，所述第二确定单元包括：
第二计算子单元，用于通过无参数相关函数计算所述候选图像序列中每一帧的第二降维特征向量与所述候选图像序列的整体特征向量的相关度，得到所述候选图像序列中每一帧的第一相关权重；
第二加权子单元，用于基于所述候选图像序列中每一帧的第一相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的自表达特征向量。
在一种可能的实现方式中，所述第一相关权重包括第一归一化相关权重，所述第一归一化相关权重是对所述第一相关权重进行归一化处理得到的。
在一种可能的实现方式中，所述第二确定模块包括：
第三确定子模块，用于将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量；
第四确定子模块，用于将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，所述第三确定子模块包括：
第三降维单元，用于通过所述第三子神经网络的第三全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第三降维特征向量；
第三确定单元，用于基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量；
所述第四确定子模块包括：
第四降维单元，用于通过所述第三子神经网络的第三全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第三降维特征向量；
第四确定单元，用于基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，所述第三确定单元包括：
第三计算子单元，用于通过无参数相关函数计算所述查询图像序列中每一帧的第三降维特征向量与所述候选图像序列的自表达特征向量的相关度，得到所述查询图像序列中每一帧的第二相关权重；
第三加权子单元，用于基于所述查询图像序列中每一帧的第二相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的协同表达特征向量。
在一种可能的实现方式中，所述第四确定单元包括：
第四计算子单元，用于通过无参数相关函数计算所述候选图像序列中每一帧的第三降维特征向量与所述查询图像序列的自表达特征向量的相关度，得到所述候选图像序列中每一帧的第二相关权重；
第四加权子单元，用于基于所述候选图像序列中每一帧的第二相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的协同表达特征向量。
在一种可能的实现方式中，所述第二相关权重包括第二归一化相关权重，所述第二归一化相关权重是对所述第二相关权重进行归一化处理得到的。
在一种可能的实现方式中，所述第三确定模块包括：
第一计算子模块，用于计算所述查询图像序列的自表达特征向量与所述候选图像序列的协同表达特征向量之差，得到第一差向量；
第二计算子模块，用于计算所述候选图像序列的自表达特征向量与所述查询图像序列的协同表达特征向量之差，得到第二差向量；
第五确定子模块，用于基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
在一种可能的实现方式中，所述第五确定子模块包括：
第一计算单元，用于计算所述第一差向量与所述第二差向量之和，得到所述查询图像序列与所述 候选图像序列的相似性特征向量；或者，
第二计算单元，用于计算所述第一差向量与所述第二差向量的相应位的元素的乘积，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
在一种可能的实现方式中，所述第四确定模块包括：
第六确定子模块，用于将所述查询图像序列与所述候选图像序列的相似性特征向量输入第四全连接层，得到所述查询图像序列与所述候选图像序列的匹配分数；
第七确定子模块，用于基于所述查询图像序列与所述候选图像序列的匹配分数，确定所述查询图像序列与所述候选图像序列的匹配结果。
在一种可能的实现方式中，所述装置还包括：
优化模块，用于基于所述查询图像序列与所述候选图像序列的匹配分数，采用同对标注数据和二元交叉熵损失函数，优化网络参数。
在一种可能的实现方式中，所述装置还包括：
第一切分模块，用于将查询视频切分为多个查询图像序列；
第二切分模块，用于将候选视频切分为多个候选图像序列；
第五确定模块，用于基于所述查询视频的查询图像序列与所述候选视频的候选图像序列的匹配结果，确定所述查询视频与所述候选视频的匹配结果。
在一种可能的实现方式中，所述第一切分模块用于：
按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，所述查询图像序列的长度等于所述预设序列长度，相邻的查询图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差；
所述第二切分模块用于：
按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，所述候选图像序列的长度等于所述预设序列长度，相邻的候选图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差。
在一种可能的实现方式中，所述第五确定模块包括：
第八确定子模块，用于确定所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数；
第三计算子模块，用于计算所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到所述查询视频与所述候选视频的匹配分数，其中，N为正整数；
第九确定子模块，用于基于所述查询视频与所述候选视频的匹配分数，确定所述查询视频与所述候选视频的匹配结果。
根据本公开的一方面，提供了一种电子设备，包括：
处理器；
用于存储处理器可执行指令的存储器；
其中，所述处理器被配置为：执行上述目标匹配方法。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述目标匹配方法。
在本公开实施例中，通过基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量，并基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果，由此能够提高目标匹配的准确性。
根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
包含在说明书中并且构成说明书的一部分的附图与说明书一起示出了本公开的示例性实施例、特征和方面，并且用于解释本公开的原理。
图1示出根据本公开实施例的目标匹配方法的流程图。
图2示出根据本公开实施例的目标匹配方法步骤S12的一示例性的流程图。
图3示出根据本公开实施例的目标匹配方法步骤S121的一示例性的流程图。
图4示出根据本公开实施例的目标匹配方法步骤S122的一示例性的流程图。
图5示出根据本公开实施例的目标匹配方法步骤S1213的一示例性的流程图。
图6示出根据本公开实施例的目标匹配方法步骤S1223的一示例性的流程图。
图7示出根据本公开实施例的目标匹配方法步骤S13的一示例性的流程图。
图8示出根据本公开实施例的目标匹配方法步骤S131的一示例性的流程图。
图9示出根据本公开实施例的目标匹配方法步骤S132的一示例性的流程图。
图10示出根据本公开实施例的目标匹配方法步骤S1312的一示例性的流程图。
图11示出根据本公开实施例的目标匹配方法步骤S1322的一示例性的流程图。
图12示出根据本公开实施例的目标匹配方法步骤S14的一示例性的流程图。
图13示出根据本公开实施例的目标匹配方法步骤S15的一示例性的流程图。
图14示出根据本公开实施例的目标匹配方法的一示例性的流程图。
图15示出根据本公开实施例的目标匹配方法步骤S28的一示例性的流程图。
图16示出根据本公开实施例的目标匹配装置的框图。
图17示出根据本公开实施例的目标匹配装置的一示例性的框图。
图18是根据一示例性实施例示出的一种电子设备800的框图。
图19是根据一示例性实施例示出的一种电子设备1900的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实 施例不必解释为优于或好于其它实施例。
另外，为了更好的说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
图1示出根据本公开实施例的目标匹配方法的流程图。本公开实施例可以应用于智能视频分析或者安防监控等领域中。例如，本公开实施例可以与行人检测、行人跟踪等技术相结合，应用于机场、车站、校园或者商场等场所的安防监控系统中。如图1所示，该方法包括步骤S11至步骤S15。
在步骤S11中，分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，查询图像序列包含待匹配目标。
在本公开实施例中，查询图像序列可以指需要进行目标匹配的图像序列。候选图像序列可以指数据库中的图像序列。数据库可以包含多个候选图像序列，例如，数据库可以包含大量的候选图像序列。在本公开实施例中，查询图像序列可以仅包含一个待匹配目标，也可以包含多个待匹配目标。本公开实施例中的图像序列可以是视频、视频片段或者其他图像序列。
在本公开实施例中，查询图像序列与候选图像序列的帧数可以不同，也可以相同。例如，查询图像序列包含T帧(即，查询图像序列包含T个图像)，候选图像序列包含R帧(即，候选图像序列包含R个图像)，其中，T和R均为正整数。
在本公开实施例中，提取查询图像序列中每一帧的特征向量，得到
Figure PCTCN2019086670-appb-000001
其中，x t表示查询图像序列中的第t帧的特征向量，1≤t≤T；提取候选图像序列中每一帧的特征向量，得到
Figure PCTCN2019086670-appb-000002
其中，y r表示候选图像序列中的第r帧的特征向量，1≤r≤R。
在一种可能的实现方式中，分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，包括：通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。例如，第一子神经网络可以为CNN(Convolutional Neural Network，卷积神经网络)。在该实现方式中，可以采用相同参数的卷积神经网络分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
在一种可能的实现方式中，在提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量之后，该方法还包括：通过第一子神经网络的第一全连接层对查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第一降维特征向量和候选图像序列中每一帧的第一降维特征向量。例如，查询图像序列中每一帧的第一降维特征向量可以表示为
Figure PCTCN2019086670-appb-000003
其中，
Figure PCTCN2019086670-appb-000004
表示查询图像序列中的第t帧的第一降维特征向量；候选图像序列中每一帧的第一降维特征向量可以表示为
Figure PCTCN2019086670-appb-000005
其中，
Figure PCTCN2019086670-appb-000006
表示候选图像序列中的第r帧的第一降维特征向量。例如，查询图像序列中每一帧的特征向量的维数为2048维，查询图像序列中每一帧的第一降维特征向量的维数为128维；候选图像序列中每一帧的特征向量的维数为2048维，候选 图像序列中每一帧的第一降维特征向量的维数为128维。例如，第一全连接层可以表示为fc-0。
在步骤S12中，分别基于查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，确定查询图像序列的自表达特征向量和候选图像序列的自表达特征向量。
在本公开实施例中，可以基于查询图像序列中每一帧的特征向量，确定查询图像序列的自表达特征向量；基于候选图像序列中每一帧的特征向量，确定候选图像序列的自表达特征向量。在本公开实施例中，查询图像序列的自表达特征向量可以表示仅由查询图像序列的表达确定的特征向量，即，查询图像序列的自表达特征向量仅由查询图像序列的表达确定，与候选图像序列的表达无关；候选图像序列的自表达特征向量可以表示仅由候选图像序列的表达确定的特征向量，即，候选图像序列的自表达特征向量仅由候选图像序列的表达确定，与查询图像序列的表达无关。
在步骤S13中，基于查询图像序列中每一帧的特征向量和候选图像序列的自表达特征向量，确定查询图像序列的协同表达特征向量，以及基于候选图像序列中每一帧的特征向量和查询图像序列的自表达特征向量，确定候选图像序列的协同表达特征向量。
在本公开实施例中，查询图像序列的协同表达特征向量可以表示由查询图像序列的表达和候选图像序列的表达共同确定的特征向量，即，查询图像序列的协同表达特征向量不仅与查询图像序列的表达相关，还与候选图像序列的表达相关；候选图像序列的协同表达特征向量可以表示由候选图像序列的表达和查询图像序列的表达共同确定的特征向量，即，候选图像序列的协同表达特征向量不仅与候选图像序列的表达相关，还与查询图像序列的表达相关。
在步骤S14中，基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量。
在本公开实施例中，查询图像序列与候选图像序列的相似性特征向量可以用于确定查询图像序列与候选图像序列的相似程度，从而可以用于判断查询图像序列与候选图像序列是否匹配。
在步骤S15中，基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果。
本公开实施例中相匹配的两个图像序列，可能是从不同的拍摄视角拍摄的同一人物的图像序列，也可能是从同一拍摄视角拍摄的同一人物的图像序列。
本公开实施例通过基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量，并基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果，由此能够提高目标匹配的准确性。
图2示出根据本公开实施例的目标匹配方法步骤S12的一示例性的流程图。如图2所示，步骤S12可以包括步骤S121和步骤S122。
在步骤S121中，将查询图像序列中每一帧的特征向量和查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定查询图像序列的自表达特征向量。
例如，第二子神经网络可以为SAN(Self Attention Subnetwork，基于注意力机制的自表达子神经网络)。
例如，可以将查询图像序列中每一帧的特征向量
Figure PCTCN2019086670-appb-000007
和查询图像序列中每一帧的第一降维特征向量
Figure PCTCN2019086670-appb-000008
输入第二子神经网络中，确定查询图像序列的自表达特征向量
Figure PCTCN2019086670-appb-000009
在步骤S122中，将候选图像序列中每一帧的特征向量和候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定候选图像序列的自表达特征向量。
例如，可以将候选图像序列中每一帧的特征向量
Figure PCTCN2019086670-appb-000010
和候选图像序列中每一帧的第一降维特征向量
Figure PCTCN2019086670-appb-000011
输入第二子神经网络中，确定候选图像序列的自表达特征向量
Figure PCTCN2019086670-appb-000012
图3示出根据本公开实施例的目标匹配方法步骤S121的一示例性的流程图。如图3所示，步骤S121可以包括步骤S1211和步骤S1213。
在步骤S1211中，通过第二子神经网络的第二全连接层对查询图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第二降维特征向量。
例如，查询图像序列中每一帧的第二降维特征向量可以表示为
Figure PCTCN2019086670-appb-000013
其中，
Figure PCTCN2019086670-appb-000014
表示查询图像序列中的第t帧的第二降维特征向量。
例如，第二全连接层可以表示为fc-1。
例如，查询图像序列中每一帧的第二降维特征向量的维数为128维。
在步骤S1212中，将查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到查询图像序列的整体特征向量。
例如，查询图像序列的整体特征向量可以表示为
Figure PCTCN2019086670-appb-000015
在步骤S1213中，基于查询图像序列中每一帧的第二降维特征向量、查询图像序列的整体特征向量以及查询图像序列中每一帧的第一降维特征向量，确定查询图像序列的自表达特征向量。
图4示出根据本公开实施例的目标匹配方法步骤S122的一示例性的流程图。如图4所示，步骤S122可以包括步骤S1221和步骤S1223。
在步骤S1221中，通过第二子神经网络的第二全连接层对候选图像序列中每一帧的特征向量进行降维处理，得到候选图像序列中每一帧的第二降维特征向量。
例如，候选图像序列中每一帧的第二降维特征向量的维数为128维。
在步骤S1222中，将候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到候选图像序列的整体特征向量。
例如，候选图像序列的整体特征向量可以表示为
Figure PCTCN2019086670-appb-000016
在步骤S1223中，基于候选图像序列中每一帧的第二降维特征向量、候选图像序列的整体特征向量以及候选图像序列中每一帧的第一降维特征向量，确定候选图像序列的自表达特征向量。
图5示出根据本公开实施例的目标匹配方法步骤S1213的一示例性的流程图。如图5所示，步骤S1213可以包括步骤S12131和步骤S12132。
在步骤S12131中，通过无参数相关函数计算查询图像序列中每一帧的第二降维特征向量与查询图 像序列的整体特征向量的相关度，得到查询图像序列中每一帧的第一相关权重。
例如，可以通过无参数相关函数f()计算查询图像序列中每一帧的第二降维特征向量与查询图像序列的整体特征向量的相关度，得到查询图像序列中每一帧的第一相关权重
Figure PCTCN2019086670-appb-000017
在一种可能的实现方式中，无参数相关函数f()可以采用点乘的方式计算
Figure PCTCN2019086670-appb-000018
与
Figure PCTCN2019086670-appb-000019
的相关度。
本公开实施例基于自表达机制，通过查询图像序列的自己的表达对查询图像序列的每一帧赋予相关权重。
在步骤S12132中，基于查询图像序列中每一帧的第一相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的自表达特征向量。
例如，查询图像序列的自表达特征向量可以表示为
Figure PCTCN2019086670-appb-000020
其中，
Figure PCTCN2019086670-appb-000021
表示查询图像序列中的第t帧的第二降维特征向量，
Figure PCTCN2019086670-appb-000022
表示查询图像序列的整体特征向量，
Figure PCTCN2019086670-appb-000023
表示查询图像序列中的第t帧的第一降维特征向量。
在一种可能的实现方式中，第一相关权重包括第一归一化相关权重，第一归一化相关权重是对第一相关权重进行归一化处理得到的。在该实现方式中，基于查询图像序列中每一帧的第一相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的自表达特征向量，包括：对查询图像序列中每一帧的第一相关权重进行归一化处理，得到查询图像序列中每一帧的第一归一化相关权重；基于查询图像序列中每一帧的第一归一化相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的自表达特征向量。在该实现方式中，可以采用softmax对查询图像序列中每一帧的第一相关权重进行归一化处理，得到查询图像序列中每一帧的第一归一化相关权重。
图6示出根据本公开实施例的目标匹配方法步骤S1223的一示例性的流程图。如图6所示，步骤S1223可以包括步骤S12231和步骤S12232。
在步骤S12231中，通过无参数相关函数计算候选图像序列中每一帧的第二降维特征向量与候选图像序列的整体特征向量的相关度，得到候选图像序列中每一帧的第一相关权重。
例如，可以通过无参数相关函数f()计算候选图像序列中每一帧的第二降维特征向量与候选图像序列的整体特征向量的相关度，得到候选图像序列中每一帧的第一相关权重
Figure PCTCN2019086670-appb-000024
在一种可能的实现方式中，无参数相关函数f()可以采用点乘的方式计算
Figure PCTCN2019086670-appb-000025
与
Figure PCTCN2019086670-appb-000026
的相关度。
本公开实施例基于自表达机制，通过候选图像序列的自己的表达对候选图像序列的每一帧赋予相关权重。
在步骤S12232中，基于候选图像序列中每一帧的第一相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的自表达特征向量。
例如，候选图像序列的自表达特征向量可以表示为
Figure PCTCN2019086670-appb-000027
其中，
Figure PCTCN2019086670-appb-000028
表示候选图像序列中的第r帧的第二降维特征向量，
Figure PCTCN2019086670-appb-000029
表示候选图像序列的整体特征向量，
Figure PCTCN2019086670-appb-000030
表示候选图像序列中的第r帧的第一降维特征向量。
在一种可能的实现方式中，第一相关权重包括第一归一化相关权重，第一归一化相关权重是对第一相关权重进行归一化处理得到的。在该实现方式中，基于候选图像序列中每一帧的第一相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的自表达特征向量，包括：对候选图像序列中每一帧的第一相关权重进行归一化处理，得到候选图像序列中每一帧的第一归一化相关权重；基于候选图像序列中每一帧的第一归一化相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的自表达特征向量。在该实现方式中，可以采用softmax对候选图像序列中每一帧的第一相关权重进行归一化处理，得到候选图像序列中每一帧的第一归一化相关权重。
图7示出根据本公开实施例的目标匹配方法步骤S13的一示例性的流程图。如图7所示，步骤S13可以包括步骤S131和步骤S132。
在步骤S131中，将查询图像序列中每一帧的特征向量、查询图像序列中每一帧的第一降维特征向量以及候选图像序列的自表达特征向量输入第三子神经网络中，得到查询图像序列的协同表达特征向量。
例如，第三子神经网络可以为CAN(Collaborative Attention Subnetwork，基于注意力机制的协同表达子神经网络)。
例如，可以将查询图像序列中每一帧的特征向量
Figure PCTCN2019086670-appb-000031
查询图像序列中每一帧的第一降维特征向量
Figure PCTCN2019086670-appb-000032
以及候选图像序列的自表达特征向量
Figure PCTCN2019086670-appb-000033
输入第三子神经网络中，得到查询图像序列的协同表达特征向量
Figure PCTCN2019086670-appb-000034
在步骤S132中，将候选图像序列中每一帧的特征向量、候选图像序列中每一帧的第一降维特征向量以及查询图像序列的自表达特征向量输入第三子神经网络中，得到候选图像序列的协同表达特征向量。
例如，可以将候选图像序列中每一帧的特征向量
Figure PCTCN2019086670-appb-000035
候选图像序列中每一帧的第一降维特征向量
Figure PCTCN2019086670-appb-000036
以及查询图像序列的自表达特征向量
Figure PCTCN2019086670-appb-000037
输入第三子神经网络中，得到候选图像序列的协同表达特征向量
Figure PCTCN2019086670-appb-000038
图8示出根据本公开实施例的目标匹配方法步骤S131的一示例性的流程图。如图8所示，步骤S131可以包括步骤S1311和步骤S1312。
在步骤S1311中，通过第三子神经网络的第三全连接层对查询图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第三降维特征向量。
例如，查询图像序列中每一帧的第三降维特征向量可以表示为
Figure PCTCN2019086670-appb-000039
其中，
Figure PCTCN2019086670-appb-000040
表示查询图像序列中第t帧的第三降维特征向量。例如，查询图像序列中每一帧的第三降维特征向量的维数为128维。
例如，第三全连接层可以表示为fc-2。
在步骤S1312中，基于查询图像序列中每一帧的第三降维特征向量、候选图像序列的自表达特征向量以及查询图像序列中每一帧的第一降维特征向量，得到查询图像序列的协同表达特征向量。
图9示出根据本公开实施例的目标匹配方法步骤S132的一示例性的流程图。如图9所示，步骤S132可以包括步骤S1321和步骤S1322。
在步骤S1321中，通过第三子神经网络的第三全连接层对候选图像序列中每一帧的特征向量进行降维处理，得到候选图像序列中每一帧的第三降维特征向量。
例如，候选图像序列中每一帧的第三降维特征向量可以表示为
Figure PCTCN2019086670-appb-000041
其中，
Figure PCTCN2019086670-appb-000042
表示候选图像序列中第r帧的第三降维特征向量。例如，候选图像序列中每一帧的第三降维特征向量的维数为128维。
在步骤S1322中，基于候选图像序列中每一帧的第三降维特征向量、查询图像序列的自表达特征向量以及候选图像序列中每一帧的第一降维特征向量，得到候选图像序列的协同表达特征向量。
图10示出根据本公开实施例的目标匹配方法步骤S1312的一示例性的流程图。如图10所示，步骤S1312可以包括步骤S13121和步骤S13122。
在步骤S13121中，通过无参数相关函数计算查询图像序列中每一帧的第三降维特征向量与候选图像序列的自表达特征向量的相关度，得到查询图像序列中每一帧的第二相关权重。
例如，查询图像序列中的第t帧的第二相关权重可以表示为
Figure PCTCN2019086670-appb-000043
本公开实施例基于协同表达机制，通过候选图像序列的表达和查询图像序列的自己的表达对查询图像序列的每一帧赋予相关权重。
在步骤S13122中，基于查询图像序列中每一帧的第二相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的协同表达特征向量。
例如，查询图像序列的协同表达特征向量可以表示为
Figure PCTCN2019086670-appb-000044
在一种可能的实现方式中，第二相关权重包括第二归一化相关权重，第二归一化相关权重是对第二相关权重进行归一化处理得到的。在该实现方式中，基于查询图像序列中每一帧的第二相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的协同表达特征向量，包括：对查询图像序列中每一帧的第二相关权重进行归一化处理，得到查询图像序列中每一帧的第二归一化相关权重；基于查询图像序列中每一帧的第二归一化相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的协同表达特征向量。
图11示出根据本公开实施例的目标匹配方法步骤S1322的一示例性的流程图。如图11所示，步骤S1322可以包括步骤S13221和步骤S13222。
在步骤S13221中，通过无参数相关函数计算候选图像序列中每一帧的第三降维特征向量与查询图像序列的自表达特征向量的相关度，得到候选图像序列中每一帧的第二相关权重。
例如，候选图像序列中第r帧的第二相关权重可以表示为
Figure PCTCN2019086670-appb-000045
本公开实施例基于协同表达机制，通过查询图像序列的表达和候选图像序列的自己的表达对候选图像序列的每一帧赋予相关权重。
在步骤S13222中，基于候选图像序列中每一帧的第二相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的协同表达特征向量。
例如，候选图像序列的协同表达特征向量可以表示为
Figure PCTCN2019086670-appb-000046
在一种可能的实现方式中，第二相关权重包括第二归一化相关权重，第二归一化相关权重是对第二相关权重进行归一化处理得到的。在该实现方式中，基于候选图像序列中每一帧的第二相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的协同表达特征向量，包括：对候选图像序列中每一帧的第二相关权重进行归一化处理，得到候选图像序列中每一帧的第二归一化相关权重；基于候选图像序列中每一帧的第二归一化相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的协同表达特征向量。
在本公开实施例中，第二子神经网络和第三子神经网络基于自表达机制和协同表达机制，通过查询图像序列的表达和候选图像序列的表达为查询图像序列的每一帧和候选图像序列的每一帧赋予相关权重。第二子神经网络和第三子神经网络利用这种非参数的自表达和协同表达，隐式地对查询图像序列和候选图像序列进行帧对齐，从而选择更具有判别性的帧对两段图像序列进行表达。由于第二子神经网络和第三子神经网络是非参数的，因此允许查询图像序列与候选图像序列具有不同的长度，因此本公开实施例提供的目标匹配方法的灵活性较高，能够广泛应用。
图12示出根据本公开实施例的目标匹配方法步骤S14的一示例性的流程图。如图12所示，步骤S14可以包括步骤S141和步骤S143。
在步骤S141中，计算查询图像序列的自表达特征向量与候选图像序列的协同表达特征向量之差，得到第一差向量。
例如，第一差向量为
Figure PCTCN2019086670-appb-000047
在步骤S142中，计算候选图像序列的自表达特征向量与查询图像序列的协同表达特征向量之差，得到第二差向量。
例如，第二差向量为
Figure PCTCN2019086670-appb-000048
在步骤S143中，基于第一差向量与第二差向量，得到查询图像序列与候选图像序列的相似性特征向量。
在一种可能的实现方式中，基于第一差向量与第二差向量，得到查询图像序列与候选图像序列的 相似性特征向量，包括：计算第一差向量与第二差向量之和，得到查询图像序列与候选图像序列的相似性特征向量。例如，查询图像序列与候选图像序列的相似性特征向量
Figure PCTCN2019086670-appb-000049
在另一种可能的实现方式中，基于第一差向量与第二差向量，得到查询图像序列与候选图像序列的相似性特征向量，包括：计算第一差向量与第二差向量的相应位的元素的乘积，得到查询图像序列与候选图像序列的相似性特征向量。
图13示出根据本公开实施例的目标匹配方法步骤S15的一示例性的流程图。如图13所示，步骤S15可以包括步骤S151和步骤S152。
在步骤S151中，将查询图像序列与候选图像序列的相似性特征向量输入第四全连接层，得到查询图像序列与候选图像序列的匹配分数。
例如，第四全连接层可以表示为fc-3。
需要说明的是，本公开实施例中不同全连接层的参数可以不相同。第一全连接层、第二全连接层、第三全连接层和第四全连接层中的“第一”“第二”“第三”和“第四”仅为表述和指代的方便，表示这四个全连接层可以是不同的全连接层。“第一”“第二”“第三”和“第四”并不用于限定全连接层的连接顺序。
在步骤S152中，基于查询图像序列与候选图像序列的匹配分数，确定查询图像序列与候选图像序列的匹配结果。
例如，若查询图像序列与候选图像序列的匹配分数大于分数阈值，则可以确定查询图像序列与候选图像序列的匹配结果为查询图像序列与候选图像序列相匹配；若查询图像序列与候选图像序列的匹配分数小于或等于分数阈值，则可以确定查询图像序列与候选图像序列的匹配结果为查询图像序列与候选图像序列不匹配。
在一种可能的实现方式中，在得到查询图像序列与候选图像序列的匹配分数之后，该方法还包括：基于查询图像序列与候选图像序列的匹配分数，采用同对标注数据和二元交叉熵损失函数，优化网络参数。
作为该实现方式的一个示例，可以采用
Figure PCTCN2019086670-appb-000050
优化网络参数。其中，N表示训练集中查询图像序列与候选图像序列对的数量，m i表示第i对的匹配分数，若第i对的查询图像序列与候选图像序列对属于同一人物，则l i＝1，否则l i＝0。
在本公开实施例中，在训练过程中，可以对训练图像序列进行切分，生成丰富的查询图像序列与候选图像序列对，从而有效地提升优化效率，并能提高网络模型的鲁棒性，从而能够提高匹配精度。
图14示出根据本公开实施例的目标匹配方法的一示例性的流程图。如图14所示，该方法可以包括步骤S21至步骤S28。
在步骤S21中，将查询视频切分为多个查询图像序列。
在一种可能的实现方式中，将查询视频切分为多个查询图像序列，包括：按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，查询图像序列的长度等于预设序列长度，相 邻的查询图像序列之间重叠的图像数等于预设序列长度与预设步长之差。
在步骤S22中，将候选视频切分为多个候选图像序列。
在一种可能的实现方式中，将候选视频切分为多个候选图像序列，包括：按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，候选图像序列的长度等于预设序列长度，相邻的候选图像序列之间重叠的图像数等于预设序列长度与预设步长之差。
在步骤S23中，分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，查询图像序列包含待匹配目标。
其中，对步骤S23参见上文对步骤S11的描述。
在步骤S24中，分别基于查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，确定查询图像序列的自表达特征向量和候选图像序列的自表达特征向量。
其中，对步骤S24参见上文对步骤S12的描述。
在步骤S25中，基于查询图像序列中每一帧的特征向量和候选图像序列的自表达特征向量，确定查询图像序列的协同表达特征向量，以及基于候选图像序列中每一帧的特征向量和查询图像序列的自表达特征向量，确定候选图像序列的协同表达特征向量。
其中，对步骤S25参见上文对步骤S13的描述。
在步骤S26中，基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量。
其中，对步骤S26参见上文对步骤S14的描述。
在步骤S27中，基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果。
其中，对步骤S27参见上文对步骤S15的描述。
在步骤S28中，基于查询视频的查询图像序列与候选视频的候选图像序列的匹配结果，确定查询视频与候选视频的匹配结果。
图15示出根据本公开实施例的目标匹配方法步骤S28的一示例性的流程图。如图15所示，步骤S28可以包括步骤S281至步骤S283。
在步骤S281中，确定查询视频的各个查询图像序列与候选视频的各个候选图像序列的匹配分数。
在步骤S282中，计算查询视频的各个查询图像序列与候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到查询视频与候选视频的匹配分数，其中，N为正整数。
在步骤S283中，基于查询视频与候选视频的匹配分数，确定查询视频与候选视频的匹配结果。
在一种可能的实现方式中，若查询视频与候选视频的匹配分数大于分数阈值，则可以确定查询视频与候选视频的匹配结果为查询视频与候选视频相匹配；若查询视频与候选视频的匹配分数小于或等于分数阈值，则可以确定查询视频与候选视频的匹配结果为查询视频与候选视频不匹配。
本公开实施例提供的目标匹配方法能够筛选出图像序列中更具有区分性的关键帧，利用多个关键帧对图像序列进行表达，由此能够提升判别能力；本公开实施例提出了更为有效的时域建模方法，捕捉连续帧的动态变化信息，提升了模型的表达能力；本公开实施例提出了更为有效的距离度量方法，减小了相同人物的特征表达之间的距离，增大了不同人物的特征表达之间的距离。本公开实施例提供 的目标匹配方法在光照条件较恶劣、遮挡较严重、视角较差或者背景干扰严重的情况下，仍然能够获得较准确的目标匹配结果。利用本公开实施例，可以帮助改进行人检测和/或行人跟踪的效果。利用本公开实施例，可以在智能视频监控中对特定行人(例如犯罪嫌疑人、失踪儿童等)更好地进行跨摄像头的搜索和追踪。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。
此外，本公开还提供了目标匹配装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种目标匹配方法，相应技术方案和描述和参见方法部分的相应记载，不再赘述。
图16示出根据本公开实施例的目标匹配装置的框图。如图16所示，该装置包括：提取模块31，用于分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，查询图像序列包含待匹配目标；第一确定模块32，用于分别基于查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，确定查询图像序列的自表达特征向量和候选图像序列的自表达特征向量；第二确定模块33，用于基于查询图像序列中每一帧的特征向量和候选图像序列的自表达特征向量，确定查询图像序列的协同表达特征向量，以及基于候选图像序列中每一帧的特征向量和查询图像序列的自表达特征向量，确定候选图像序列的协同表达特征向量；第三确定模块34，用于基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量；第四确定模块35，用于基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果。
在一种可能的实现方式中，提取模块31用于：通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
图17示出根据本公开实施例的目标匹配装置的一示例性的框图。如图17所示：
在一种可能的实现方式中，该装置还包括：降维模块36，用于通过第一子神经网络的第一全连接层对查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第一降维特征向量和候选图像序列中每一帧的第一降维特征向量。
在一种可能的实现方式中，第一确定模块32包括：第一确定子模块321，用于将查询图像序列中每一帧的特征向量和查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定查询图像序列的自表达特征向量；第二确定子模块322，用于将候选图像序列中每一帧的特征向量和候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定候选图像序列的自表达特征向量。
在一种可能的实现方式中，第一确定子模块321包括：第一降维单元，用于通过第二子神经网络的第二全连接层对查询图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第二降维特征向量；第一平均池化单元，用于将查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到查询图像序列的整体特征向量；第一确定单元，用于基于查询图像序列中每一帧的第二降维特征向量、查询图像序列的整体特征向量以及查询图像序列中每一帧的第一降维特征向量，确定查询图像序列的自表达特征向量。
在一种可能的实现方式中，第二确定子模块322包括：第二降维单元，用于通过第二子神经网络的第二全连接层对候选图像序列中每一帧的特征向量进行降维处理，得到候选图像序列中每一帧的第 二降维特征向量；第二平均池化单元，用于将候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到候选图像序列的整体特征向量；第二确定单元，用于基于候选图像序列中每一帧的第二降维特征向量、候选图像序列的整体特征向量以及候选图像序列中每一帧的第一降维特征向量，确定候选图像序列的自表达特征向量。
在一种可能的实现方式中，第一确定单元包括：第一计算子单元，用于通过无参数相关函数计算查询图像序列中每一帧的第二降维特征向量与查询图像序列的整体特征向量的相关度，得到查询图像序列中每一帧的第一相关权重；第一加权子单元，用于基于查询图像序列中每一帧的第一相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的自表达特征向量。
在一种可能的实现方式中，第二确定单元包括：第二计算子单元，用于通过无参数相关函数计算候选图像序列中每一帧的第二降维特征向量与候选图像序列的整体特征向量的相关度，得到候选图像序列中每一帧的第一相关权重；第二加权子单元，用于基于候选图像序列中每一帧的第一相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的自表达特征向量。
在一种可能的实现方式中，第一相关权重包括第一归一化相关权重，第一归一化相关权重是对第一相关权重进行归一化处理得到的。
在一种可能的实现方式中，第二确定模块33包括：第三确定子模块331，用于将查询图像序列中每一帧的特征向量、查询图像序列中每一帧的第一降维特征向量以及候选图像序列的自表达特征向量输入第三子神经网络中，得到查询图像序列的协同表达特征向量；第四确定子模块332，用于将候选图像序列中每一帧的特征向量、候选图像序列中每一帧的第一降维特征向量以及查询图像序列的自表达特征向量输入第三子神经网络中，得到候选图像序列的协同表达特征向量。
在一种可能的实现方式中，第三确定子模块331包括：第三降维单元，用于通过第三子神经网络的第三全连接层对查询图像序列中每一帧的特征向量进行降维处理，得到查询图像序列中每一帧的第三降维特征向量；第三确定单元，用于基于查询图像序列中每一帧的第三降维特征向量、候选图像序列的自表达特征向量以及查询图像序列中每一帧的第一降维特征向量，得到查询图像序列的协同表达特征向量；第四确定子模块332包括：第四降维单元，用于通过第三子神经网络的第三全连接层对候选图像序列中每一帧的特征向量进行降维处理，得到候选图像序列中每一帧的第三降维特征向量；第四确定单元，用于基于候选图像序列中每一帧的第三降维特征向量、查询图像序列的自表达特征向量以及候选图像序列中每一帧的第一降维特征向量，得到候选图像序列的协同表达特征向量。
在一种可能的实现方式中，第三确定单元包括：第三计算子单元，用于通过无参数相关函数计算查询图像序列中每一帧的第三降维特征向量与候选图像序列的自表达特征向量的相关度，得到查询图像序列中每一帧的第二相关权重；第三加权子单元，用于基于查询图像序列中每一帧的第二相关权重，对查询图像序列中每一帧的第一降维特征向量进行加权，得到查询图像序列的协同表达特征向量。
在一种可能的实现方式中，第四确定单元包括：第四计算子单元，用于通过无参数相关函数计算候选图像序列中每一帧的第三降维特征向量与查询图像序列的自表达特征向量的相关度，得到候选图像序列中每一帧的第二相关权重；第四加权子单元，用于基于候选图像序列中每一帧的第二相关权重，对候选图像序列中每一帧的第一降维特征向量进行加权，得到候选图像序列的协同表达特征向量。
在一种可能的实现方式中，第二相关权重包括第二归一化相关权重，第二归一化相关权重是对第 二相关权重进行归一化处理得到的。
在一种可能的实现方式中，第三确定模块34包括：第一计算子模块341，用于计算查询图像序列的自表达特征向量与候选图像序列的协同表达特征向量之差，得到第一差向量；第二计算子模块342，用于计算候选图像序列的自表达特征向量与查询图像序列的协同表达特征向量之差，得到第二差向量；第五确定子模块343，用于基于第一差向量与第二差向量，得到查询图像序列与候选图像序列的相似性特征向量。
在一种可能的实现方式中，第五确定子模块343包括：第一计算单元，用于计算第一差向量与第二差向量之和，得到查询图像序列与候选图像序列的相似性特征向量；或者，第二计算单元，用于计算第一差向量与第二差向量的相应位的元素的乘积，得到查询图像序列与候选图像序列的相似性特征向量。
在一种可能的实现方式中，第四确定模块35包括：第六确定子模块351，用于将查询图像序列与候选图像序列的相似性特征向量输入第四全连接层，得到查询图像序列与候选图像序列的匹配分数；第七确定子模块352，用于基于查询图像序列与候选图像序列的匹配分数，确定查询图像序列与候选图像序列的匹配结果。
在一种可能的实现方式中，该装置还包括：优化模块37，用于基于查询图像序列与候选图像序列的匹配分数，采用同对标注数据和二元交叉熵损失函数，优化网络参数。
在一种可能的实现方式中，该装置还包括：第一切分模块38，用于将查询视频切分为多个查询图像序列；第二切分模块39，用于将候选视频切分为多个候选图像序列；第五确定模块30，用于基于查询视频的查询图像序列与候选视频的候选图像序列的匹配结果，确定查询视频与候选视频的匹配结果。
在一种可能的实现方式中，第一切分模块38用于：按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，查询图像序列的长度等于预设序列长度，相邻的查询图像序列之间重叠的图像数等于预设序列长度与预设步长之差；第二切分模块39用于：按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，候选图像序列的长度等于预设序列长度，相邻的候选图像序列之间重叠的图像数等于预设序列长度与预设步长之差。
在一种可能的实现方式中，第五确定模块30包括：第八确定子模块301，用于确定查询视频的各个查询图像序列与候选视频的各个候选图像序列的匹配分数；第三计算子模块302，用于计算查询视频的各个查询图像序列与候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到查询视频与候选视频的匹配分数，其中，N为正整数；第九确定子模块303，用于基于查询视频与候选视频的匹配分数，确定查询视频与候选视频的匹配结果。
本公开实施例通过基于查询图像序列的自表达特征向量、查询图像序列的协同表达特征向量、候选图像序列的自表达特征向量以及候选图像序列的协同表达特征向量，确定查询图像序列与候选图像序列的相似性特征向量，并基于相似性特征向量，确定查询图像序列与候选图像序列的匹配结果，由此能够提高目标匹配的准确性。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中， 所述处理器被配置为上述方法。
电子设备可以被提供为终端、服务器或其它形态的设备。
图18是根据一示例性实施例示出的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图18，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设备800方位或加速/减速和电子设备800的温度变化。 传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如CMOS或CCD图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如WiFi，2G或3G，或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
图19是根据一示例性实施例示出的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图19，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是――但不限于――电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光 脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合 来实现。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (44)

  1. 一种目标匹配方法，其特征在于，包括：
    分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，所述查询图像序列包含待匹配目标；
    分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量；
    基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量；
    基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，确定所述查询图像序列与所述候选图像序列的相似性特征向量；
    基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果。
  2. 根据权利要求1所述的方法，其特征在于，分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，包括：
    通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
  3. 根据权利要求1或2所述的方法，其特征在于，在提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量之后，所述方法还包括：
    通过第一子神经网络的第一全连接层对所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第一降维特征向量和所述候选图像序列中每一帧的第一降维特征向量。
  4. 根据权利要求3所述的方法，其特征在于，分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量包括：
    将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量；
    将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述候选图像序列的自表达特征向量。
  5. 根据权利要求4所述的方法，其特征在于，将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量，包括：
    通过所述第二子神经网络的第二全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第二降维特征向量；
    将所述查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述查询图像序列的整体特征向量；
    基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的整体特征向量以及所 述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量。
  6. 根据权利要求4所述的方法，其特征在于，将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，得到所述候选图像序列的自表达特征向量，包括：
    通过所述第二子神经网络的第二全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第二降维特征向量；
    将所述候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述候选图像序列的整体特征向量；
    基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量。
  7. 根据权利要求5所述的方法，其特征在于，基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的整体特征向量以及所述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量，包括：
    通过无参数相关函数计算所述查询图像序列中每一帧的第二降维特征向量与所述查询图像序列的整体特征向量的相关度，得到所述查询图像序列中每一帧的第一相关权重；
    基于所述查询图像序列中每一帧的第一相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的自表达特征向量。
  8. 根据权利要求6所述的方法，其特征在于，基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量，包括：
    通过无参数相关函数计算所述候选图像序列中每一帧的第二降维特征向量与所述候选图像序列的整体特征向量的相关度，得到所述候选图像序列中每一帧的第一相关权重；
    基于所述候选图像序列中每一帧的第一相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的自表达特征向量。
  9. 根据权利要求7或8所述的方法，其特征在于，所述第一相关权重包括第一归一化相关权重，所述第一归一化相关权重是对所述第一相关权重进行归一化处理得到的。
  10. 根据权利要求3至9中任意一项所述的方法，其特征在于，基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量，包括：
    将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量；
    将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量。
  11. 根据权利要求10所述的方法，其特征在于，将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量，包括：
    通过所述第三子神经网络的第三全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第三降维特征向量；
    基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量；
    将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量，包括：
    通过所述第三子神经网络的第三全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第三降维特征向量；
    基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量。
  12. 根据权利要求11所述的方法，其特征在于，基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量，包括：
    通过无参数相关函数计算所述查询图像序列中每一帧的第三降维特征向量与所述候选图像序列的自表达特征向量的相关度，得到所述查询图像序列中每一帧的第二相关权重；
    基于所述查询图像序列中每一帧的第二相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的协同表达特征向量。
  13. 根据权利要求11所述的方法，其特征在于，基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量，包括：
    通过无参数相关函数计算所述候选图像序列中每一帧的第三降维特征向量与所述查询图像序列的自表达特征向量的相关度，得到所述候选图像序列中每一帧的第二相关权重；
    基于所述候选图像序列中每一帧的第二相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的协同表达特征向量。
  14. 根据权利要求12或13所述的方法，其特征在于，所述第二相关权重包括第二归一化相关权重，所述第二归一化相关权重是对所述第二相关权重进行归一化处理得到的。
  15. 根据权利要求1至14中任意一项所述的方法，其特征在于，基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量，包括：
    计算所述查询图像序列的自表达特征向量与所述候选图像序列的协同表达特征向量之差，得到第一差向量；
    计算所述候选图像序列的自表达特征向量与所述查询图像序列的协同表达特征向量之差，得到第二差向量；
    基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
  16. 根据权利要求15所述的方法，其特征在于，基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量，包括：
    计算所述第一差向量与所述第二差向量之和，得到所述查询图像序列与所述候选图像序列的相似性特征向量；或者，
    计算所述第一差向量与所述第二差向量的相应位的元素的乘积，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
  17. 根据权利要求1至16中任意一项所述的方法，其特征在于，基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果，包括：
    将所述查询图像序列与所述候选图像序列的相似性特征向量输入第四全连接层，得到所述查询图像序列与所述候选图像序列的匹配分数；
    基于所述查询图像序列与所述候选图像序列的匹配分数，确定所述查询图像序列与所述候选图像序列的匹配结果。
  18. 根据权利要求17所述的方法，其特征在于，在得到所述查询图像序列与所述候选图像序列的匹配分数之后，所述方法还包括：
    基于所述查询图像序列与所述候选图像序列的匹配分数，采用同对标注数据和二元交叉熵损失函数，优化网络参数。
  19. 根据权利要求1至18中任意一项所述的方法，其特征在于，在提取查询图像序列中每一帧的特征向量之前，所述方法还包括：
    将查询视频切分为多个查询图像序列；
    将候选视频切分为多个候选图像序列；
    在确定所述查询图像序列与所述候选图像序列的匹配结果之后，所述方法还包括：
    基于所述查询视频的查询图像序列与所述候选视频的候选图像序列的匹配结果，确定所述查询视频与所述候选视频的匹配结果。
  20. 根据权利要求19所述的方法，其特征在于，将查询视频切分为多个查询图像序列，包括：
    按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，所述查询图像序列的长度等于所述预设序列长度，相邻的查询图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差；
    将候选视频切分为多个候选图像序列，包括：
    按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，所述候选图像序列的长度等于所述预设序列长度，相邻的候选图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差。
  21. 根据权利要求19或20所述的方法，其特征在于，基于所述查询视频的查询图像序列与所述候 选视频的候选图像序列的匹配结果，确定所述查询视频与所述候选视频的匹配结果，包括：
    确定所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数；
    计算所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到所述查询视频与所述候选视频的匹配分数，其中，N为正整数；
    基于所述查询视频与所述候选视频的匹配分数，确定所述查询视频与所述候选视频的匹配结果。
  22. 一种目标匹配装置，其特征在于，包括：
    提取模块，用于分别提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量，其中，所述查询图像序列包含待匹配目标；
    第一确定模块，用于分别基于所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量，确定所述查询图像序列的自表达特征向量和所述候选图像序列的自表达特征向量；
    第二确定模块，用于基于所述查询图像序列中每一帧的特征向量和所述候选图像序列的自表达特征向量，确定所述查询图像序列的协同表达特征向量，以及基于所述候选图像序列中每一帧的特征向量和所述查询图像序列的自表达特征向量，确定所述候选图像序列的协同表达特征向量；
    第三确定模块，用于基于所述查询图像序列的自表达特征向量、所述查询图像序列的协同表达特征向量、所述候选图像序列的自表达特征向量以及所述候选图像序列的协同表达特征向量，确定所述查询图像序列与所述候选图像序列的相似性特征向量；
    第四确定模块，用于基于所述相似性特征向量，确定所述查询图像序列与所述候选图像序列的匹配结果。
  23. 根据权利要求22所述的装置，其特征在于，所述提取模块用于：
    通过第一子神经网络提取查询图像序列中每一帧的特征向量和候选图像序列中每一帧的特征向量。
  24. 根据权利要求22或23所述的装置，其特征在于，所述装置还包括：
    降维模块，用于通过第一子神经网络的第一全连接层对所述查询图像序列中每一帧的特征向量和所述候选图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第一降维特征向量和所述候选图像序列中每一帧的第一降维特征向量。
  25. 根据权利要求24所述的装置，其特征在于，所述第一确定模块包括：
    第一确定子模块，用于将所述查询图像序列中每一帧的特征向量和所述查询图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述查询图像序列的自表达特征向量；
    第二确定子模块，用于将所述候选图像序列中每一帧的特征向量和所述候选图像序列中每一帧的第一降维特征向量输入第二子神经网络中，确定所述候选图像序列的自表达特征向量。
  26. 根据权利要求25所述的装置，其特征在于，所述第一确定子模块包括：
    第一降维单元，用于通过所述第二子神经网络的第二全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第二降维特征向量；
    第一平均池化单元，用于将所述查询图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述查询图像序列的整体特征向量；
    第一确定单元，用于基于所述查询图像序列中每一帧的第二降维特征向量、所述查询图像序列的 整体特征向量以及所述查询图像序列中每一帧的第一降维特征向量，确定所述查询图像序列的自表达特征向量。
  27. 根据权利要求25所述的装置，其特征在于，所述第二确定子模块包括：
    第二降维单元，用于通过所述第二子神经网络的第二全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第二降维特征向量；
    第二平均池化单元，用于将所述候选图像序列中每一帧的第二降维特征向量经过时间维度的平均池化处理，得到所述候选图像序列的整体特征向量；
    第二确定单元，用于基于所述候选图像序列中每一帧的第二降维特征向量、所述候选图像序列的整体特征向量以及所述候选图像序列中每一帧的第一降维特征向量，确定所述候选图像序列的自表达特征向量。
  28. 根据权利要求26所述的装置，其特征在于，所述第一确定单元包括：
    第一计算子单元，用于通过无参数相关函数计算所述查询图像序列中每一帧的第二降维特征向量与所述查询图像序列的整体特征向量的相关度，得到所述查询图像序列中每一帧的第一相关权重；
    第一加权子单元，用于基于所述查询图像序列中每一帧的第一相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的自表达特征向量。
  29. 根据权利要求27所述的装置，其特征在于，所述第二确定单元包括：
    第二计算子单元，用于通过无参数相关函数计算所述候选图像序列中每一帧的第二降维特征向量与所述候选图像序列的整体特征向量的相关度，得到所述候选图像序列中每一帧的第一相关权重；
    第二加权子单元，用于基于所述候选图像序列中每一帧的第一相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的自表达特征向量。
  30. 根据权利要求28或29所述的装置，其特征在于，所述第一相关权重包括第一归一化相关权重，所述第一归一化相关权重是对所述第一相关权重进行归一化处理得到的。
  31. 根据权利要求24至30中任意一项所述的装置，其特征在于，所述第二确定模块包括：
    第三确定子模块，用于将所述查询图像序列中每一帧的特征向量、所述查询图像序列中每一帧的第一降维特征向量以及所述候选图像序列的自表达特征向量输入第三子神经网络中，得到所述查询图像序列的协同表达特征向量；
    第四确定子模块，用于将所述候选图像序列中每一帧的特征向量、所述候选图像序列中每一帧的第一降维特征向量以及所述查询图像序列的自表达特征向量输入第三子神经网络中，得到所述候选图像序列的协同表达特征向量。
  32. 根据权利要求31所述的装置，其特征在于，所述第三确定子模块包括：
    第三降维单元，用于通过所述第三子神经网络的第三全连接层对所述查询图像序列中每一帧的特征向量进行降维处理，得到所述查询图像序列中每一帧的第三降维特征向量；
    第三确定单元，用于基于所述查询图像序列中每一帧的第三降维特征向量、所述候选图像序列的自表达特征向量以及所述查询图像序列中每一帧的第一降维特征向量，得到所述查询图像序列的协同表达特征向量；
    所述第四确定子模块包括：
    第四降维单元，用于通过所述第三子神经网络的第三全连接层对所述候选图像序列中每一帧的特征向量进行降维处理，得到所述候选图像序列中每一帧的第三降维特征向量；
    第四确定单元，用于基于所述候选图像序列中每一帧的第三降维特征向量、所述查询图像序列的自表达特征向量以及所述候选图像序列中每一帧的第一降维特征向量，得到所述候选图像序列的协同表达特征向量。
  33. 根据权利要求32所述的装置，其特征在于，所述第三确定单元包括：
    第三计算子单元，用于通过无参数相关函数计算所述查询图像序列中每一帧的第三降维特征向量与所述候选图像序列的自表达特征向量的相关度，得到所述查询图像序列中每一帧的第二相关权重；
    第三加权子单元，用于基于所述查询图像序列中每一帧的第二相关权重，对所述查询图像序列中每一帧的第一降维特征向量进行加权，得到所述查询图像序列的协同表达特征向量。
  34. 根据权利要求32所述的装置，其特征在于，所述第四确定单元包括：
    第四计算子单元，用于通过无参数相关函数计算所述候选图像序列中每一帧的第三降维特征向量与所述查询图像序列的自表达特征向量的相关度，得到所述候选图像序列中每一帧的第二相关权重；
    第四加权子单元，用于基于所述候选图像序列中每一帧的第二相关权重，对所述候选图像序列中每一帧的第一降维特征向量进行加权，得到所述候选图像序列的协同表达特征向量。
  35. 根据权利要求33或34所述的装置，其特征在于，所述第二相关权重包括第二归一化相关权重，所述第二归一化相关权重是对所述第二相关权重进行归一化处理得到的。
  36. 根据权利要求22至35中任意一项所述的装置，其特征在于，所述第三确定模块包括：
    第一计算子模块，用于计算所述查询图像序列的自表达特征向量与所述候选图像序列的协同表达特征向量之差，得到第一差向量；
    第二计算子模块，用于计算所述候选图像序列的自表达特征向量与所述查询图像序列的协同表达特征向量之差，得到第二差向量；
    第五确定子模块，用于基于所述第一差向量与所述第二差向量，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
  37. 根据权利要求36所述的装置，其特征在于，所述第五确定子模块包括：
    第一计算单元，用于计算所述第一差向量与所述第二差向量之和，得到所述查询图像序列与所述候选图像序列的相似性特征向量；或者，
    第二计算单元，用于计算所述第一差向量与所述第二差向量的相应位的元素的乘积，得到所述查询图像序列与所述候选图像序列的相似性特征向量。
  38. 根据权利要求22至37中任意一项所述的装置，其特征在于，所述第四确定模块包括：
    第六确定子模块，用于将所述查询图像序列与所述候选图像序列的相似性特征向量输入第四全连接层，得到所述查询图像序列与所述候选图像序列的匹配分数；
    第七确定子模块，用于基于所述查询图像序列与所述候选图像序列的匹配分数，确定所述查询图像序列与所述候选图像序列的匹配结果。
  39. 根据权利要求38所述的装置，其特征在于，所述装置还包括：
    优化模块，用于基于所述查询图像序列与所述候选图像序列的匹配分数，采用同对标注数据和二 元交叉熵损失函数，优化网络参数。
  40. 根据权利要求22至39中任意一项所述的装置，其特征在于，所述装置还包括：
    第一切分模块，用于将查询视频切分为多个查询图像序列；
    第二切分模块，用于将候选视频切分为多个候选图像序列；
    第五确定模块，用于基于所述查询视频的查询图像序列与所述候选视频的候选图像序列的匹配结果，确定所述查询视频与所述候选视频的匹配结果。
  41. 根据权利要求40所述的装置，其特征在于，所述第一切分模块用于：
    按照预设序列长度以及预设步长，将查询视频切分为多个查询图像序列，其中，所述查询图像序列的长度等于所述预设序列长度，相邻的查询图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差；
    所述第二切分模块用于：
    按照预设序列长度以及预设步长，将候选视频切分为多个候选图像序列，其中，所述候选图像序列的长度等于所述预设序列长度，相邻的候选图像序列之间重叠的图像数等于所述预设序列长度与所述预设步长之差。
  42. 根据权利要求40或41所述的装置，其特征在于，所述第五确定模块包括：
    第八确定子模块，用于确定所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数；
    第三计算子模块，用于计算所述查询视频的各个查询图像序列与所述候选视频的各个候选图像序列的匹配分数中最高的N个匹配分数的平均值，得到所述查询视频与所述候选视频的匹配分数，其中，N为正整数；
    第九确定子模块，用于基于所述查询视频与所述候选视频的匹配分数，确定所述查询视频与所述候选视频的匹配结果。
  43. 一种电子设备，其特征在于，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为：执行权利要求1至21中任意一项所述的方法。
  44. 一种计算机可读存储介质，其上存储有计算机程序指令，其特征在于，所述计算机程序指令被处理器执行时实现权利要求1至21中任意一项所述的方法。
PCT/CN2019/086670 2018-06-15 2019-05-13 目标匹配方法及装置、电子设备和存储介质 WO2019237870A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
KR1020207007917A KR20200042513A (ko) 2018-06-15 2019-05-13 타겟 매칭 방법 및 장치, 전자 기기 및 저장 매체
SG11202003581YA SG11202003581YA (en) 2018-06-15 2019-05-13 Target matching method and apparatus, electronic device, and storage medium
JP2020515878A JP6883710B2 (ja) 2018-06-15 2019-05-13 ターゲットのマッチング方法及び装置、電子機器並びに記憶媒体
US16/841,723 US11222231B2 (en) 2018-06-15 2020-04-07 Target matching method and apparatus, electronic device, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810621959.5 2018-06-15
CN201810621959.5A CN109145150B (zh) 2018-06-15 2018-06-15 目标匹配方法及装置、电子设备和存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/841,723 Continuation US11222231B2 (en) 2018-06-15 2020-04-07 Target matching method and apparatus, electronic device, and storage medium

Publications (1)

Publication Number Publication Date
WO2019237870A1 true WO2019237870A1 (zh) 2019-12-19

Family

ID=64802036

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/086670 WO2019237870A1 (zh) 2018-06-15 2019-05-13 目标匹配方法及装置、电子设备和存储介质

Country Status (6)

Country Link
US (1) US11222231B2 (zh)
JP (1) JP6883710B2 (zh)
KR (1) KR20200042513A (zh)
CN (1) CN109145150B (zh)
SG (1) SG11202003581YA (zh)
WO (1) WO2019237870A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109145150B (zh) 2018-06-15 2021-02-12 深圳市商汤科技有限公司 目标匹配方法及装置、电子设备和存储介质
CN111435432B (zh) * 2019-01-15 2023-05-26 北京市商汤科技开发有限公司 网络优化方法及装置、图像处理方法及装置、存储介质
CN110705590B (zh) * 2019-09-02 2021-03-12 创新先进技术有限公司 通过计算机执行的、用于识别车辆部件的方法及装置
CN110866509B (zh) 2019-11-20 2023-04-28 腾讯科技（深圳）有限公司 动作识别方法、装置、计算机存储介质和计算机设备
KR102475177B1 (ko) * 2020-11-04 2022-12-07 한국전자기술연구원 영상 처리 방법 및 장치
CN113243886B (zh) * 2021-06-11 2021-11-09 四川翼飞视科技有限公司 一种基于深度学习的视力检测系统、方法和存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608234A (zh) * 2016-03-18 2016-05-25 北京京东尚科信息技术有限公司 图像检索方法和装置
CN107193983A (zh) * 2017-05-27 2017-09-22 北京小米移动软件有限公司 图像搜索方法及装置
CN107451156A (zh) * 2016-05-31 2017-12-08 杭州华为企业通信技术有限公司 一种图像再识别方法及识别装置
CN109145150A (zh) * 2018-06-15 2019-01-04 深圳市商汤科技有限公司 目标匹配方法及装置、电子设备和存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7921036B1 (en) * 2002-04-30 2011-04-05 Videomining Corporation Method and system for dynamically targeting content based on automatic demographics and behavior analysis
US10346726B2 (en) * 2014-12-15 2019-07-09 Samsung Electronics Co., Ltd. Image recognition method and apparatus, image verification method and apparatus, learning method and apparatus to recognize image, and learning method and apparatus to verify image
CN106326288B (zh) * 2015-06-30 2019-12-03 阿里巴巴集团控股有限公司 图像搜索方法及装置
US9818126B1 (en) * 2016-04-20 2017-11-14 Deep Labs Inc. Systems and methods for sensor data analysis through machine learning
CN106649663B (zh) * 2016-12-14 2018-10-16 大连理工大学 一种基于紧凑视频表征的视频拷贝检测方法
CN106886599B (zh) * 2017-02-28 2020-03-03 北京京东尚科信息技术有限公司 图像检索方法以及装置
CN107862331A (zh) * 2017-10-31 2018-03-30 华中科技大学 一种基于时间序列及cnn的不安全行为识别方法及系统
US20210264496A1 (en) * 2020-02-21 2021-08-26 Goldenspear Llc Machine learning for rapid analysis of image data via curated customer personas

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105608234A (zh) * 2016-03-18 2016-05-25 北京京东尚科信息技术有限公司 图像检索方法和装置
CN107451156A (zh) * 2016-05-31 2017-12-08 杭州华为企业通信技术有限公司 一种图像再识别方法及识别装置
CN107193983A (zh) * 2017-05-27 2017-09-22 北京小米移动软件有限公司 图像搜索方法及装置
CN109145150A (zh) * 2018-06-15 2019-01-04 深圳市商汤科技有限公司 目标匹配方法及装置、电子设备和存储介质

Also Published As

Publication number Publication date
SG11202003581YA (en) 2020-05-28
JP2020534606A (ja) 2020-11-26
US11222231B2 (en) 2022-01-11
CN109145150B (zh) 2021-02-12
CN109145150A (zh) 2019-01-04
US20200234078A1 (en) 2020-07-23
JP6883710B2 (ja) 2021-06-09
KR20200042513A (ko) 2020-04-23

Similar Documents

Publication Publication Date Title
WO2019237870A1 (zh) 目标匹配方法及装置、电子设备和存储介质
TWI759722B (zh) 神經網路訓練方法及裝置、圖像處理方法及裝置、電子設備和計算機可讀存儲介質
US11120078B2 (en) Method and device for video processing, electronic device, and storage medium
TWI766286B (zh) 圖像處理方法及圖像處理裝置、電子設備和電腦可讀儲存媒介
TWI747325B (zh) 目標對象匹配方法及目標對象匹配裝置、電子設備和電腦可讀儲存媒介
WO2021196401A1 (zh) 图像重建方法及装置、电子设备和存储介质
WO2021056808A1 (zh) 图像处理方法及装置、电子设备和存储介质
JP7110412B2 (ja) 生体検出方法及び装置、電子機器並びに記憶媒体
TWI738172B (zh) 影片處理方法及裝置、電子設備、儲存媒體和電腦程式
WO2021093375A1 (zh) 检测同行人的方法及装置、系统、电子设备和存储介质
TWI757668B (zh) 網路優化方法及裝置、圖像處理方法及裝置、儲存媒體
WO2021036382A1 (zh) 图像处理方法及装置、电子设备和存储介质
WO2021208666A1 (zh) 字符识别方法及装置、电子设备和存储介质
TW202109314A (zh) 圖像處理方法及圖像處理裝置、電子設備和電腦可讀儲存媒體
WO2020192113A1 (zh) 图像处理方法及装置、电子设备和存储介质
CN111582383B (zh) 属性识别方法及装置、电子设备和存储介质
WO2023115911A1 (zh) 对象重识别方法及装置、电子设备、存储介质和计算机程序产品
CN111259967A (zh) 图像分类及神经网络训练方法、装置、设备及存储介质
WO2016188065A1 (zh) 云名片推荐方法及装置
WO2021164100A1 (zh) 图像处理方法及装置、电子设备和存储介质
TWI770531B (zh) 人臉識別方法、電子設備和儲存介質
CN110110742B (zh) 多特征融合方法、装置、电子设备及存储介质
CN115422932A (zh) 一种词向量训练方法及装置、电子设备和存储介质
CN114842404A (zh) 时序动作提名的生成方法及装置、电子设备和存储介质
CN113807369A (zh) 目标重识别方法及装置、电子设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19818786

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020515878

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20207007917

Country of ref document: KR

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 23/04/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19818786

Country of ref document: EP

Kind code of ref document: A1