WO2019020103A1 - 目标识别方法、装置、存储介质和电子设备 - Google Patents

目标识别方法、装置、存储介质和电子设备 Download PDF

Info

Publication number
WO2019020103A1
WO2019020103A1 PCT/CN2018/097374 CN2018097374W WO2019020103A1 WO 2019020103 A1 WO2019020103 A1 WO 2019020103A1 CN 2018097374 W CN2018097374 W CN 2018097374W WO 2019020103 A1 WO2019020103 A1 WO 2019020103A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
determined
target
path
information
Prior art date
Application number
PCT/CN2018/097374
Other languages
English (en)
French (fr)
Inventor
沈岩涛
肖桐
李鸿升
伊帅
王晓刚
Original Assignee
北京市商汤科技开发有限公司
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 北京市商汤科技开发有限公司 filed Critical 北京市商汤科技开发有限公司
Priority to SG11201911625YA priority Critical patent/SG11201911625YA/en
Priority to JP2019557616A priority patent/JP6893564B2/ja
Priority to KR1020197031657A priority patent/KR102339323B1/ko
Publication of WO2019020103A1 publication Critical patent/WO2019020103A1/zh
Priority to US16/565,069 priority patent/US11200682B2/en
Priority to US17/452,776 priority patent/US20220051417A1/en
Priority to US17/453,487 priority patent/US20220058812A1/en

Links

Images

Classifications

    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
    • G06V20/47—Detecting features for summarising video content
    • 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
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G06F18/295—Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/143—Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • 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
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/48—Matching video sequences
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G06V20/54—Surveillance or monitoring of activities, e.g. for recognising suspicious objects of traffic, e.g. cars on the road, trains or boats
    • 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
    • G06N3/047—Probabilistic or stochastic networks

Definitions

  • Embodiments of the present invention relate to the field of artificial intelligence technologies, and in particular, to a target identification method, apparatus, storage medium, and electronic device.
  • Re-identification of vehicles is an important content in the field of computer vision and public safety. It has important application value in many aspects such as vehicle detection and tracking, route estimation and abnormal behavior detection.
  • vehicle re-identification techniques are based on the appearance information of the vehicle. Different from pedestrian re-identification, the difficulty in re-identifying the vehicle by simply using the vehicle appearance information lies in the appearance of many vehicles such as vehicles (such as color, Model, shape, etc.) are very similar. Especially in different vehicles of the same brand, the difference will be even smaller. Detecting and recognizing the unique details of vehicle identification information such as vehicle license plate information and vehicle interior decorations such as interior decorations may be caused by factors such as poor monitoring lens angle, poor lighting conditions, and lens blur. The robustness of detection and recognition is degraded, resulting in inaccurate detection and recognition results.
  • the embodiment of the invention provides a technical solution for target recognition.
  • a target recognition method includes: acquiring a first image and a second image, wherein the first image and the second image respectively include a target to be determined; and generating a predicted path based on the first image and the second image, The two ends of the prediction path respectively correspond to the first image and the second image; perform validity determination on the prediction path, and determine, in the first image and the second image, to be determined based on the determination result Whether the target is the same target to be determined.
  • the target to be determined is a vehicle.
  • the method before the generating the prediction path based on the first image and the second image, the method further includes: according to time information, spatial information, image feature information of the first image, and Determining, by the time information of the second image, the spatial information, the image feature information, a preliminary identical probability value of the target to be determined respectively included in the first image and the second image; And generating, by the second image, the prediction path, including: when the preliminary same probability value is greater than a preset value, generating the prediction path based on the first image and the second image.
  • the difference between the difference and the spatial information is input to a Siamese Convolutional Neural Network (Siamese-CNN) to obtain preliminary identical probability values of the target to be determined in the first image and the second image.
  • Siamese Convolutional Neural Network Siamese Convolutional Neural Network
  • the generating a prediction path based on the first image and the second image including: according to feature information of the first image, time information of the first image, the first The spatial information of the image, the feature information of the second image, the time information of the second image, and the spatial information of the second image are used to generate a predicted path of the target to be determined by a probability model.
  • the generating, by using the probability model, the predicted path of the target to be determined comprising: determining, by using a chained Markov Random Field (MRF), the acquired image set to include the Generating the information of the target, and all the images having the spatio-temporal sequence relationship with the first image and the second image; generating the prediction of the target to be determined according to the determined time information and spatial information corresponding to all the images path.
  • MRF chained Markov Random Field
  • the generating the predicted path of the target to be determined according to the determined time information and the spatial information of all the images including: generating time based on the determined time information and spatial information of all the images.
  • the first image is a first node and the second image is a predicted path of the tail node, wherein the predicted path corresponds to at least one intermediate node in addition to the first node and the tail node.
  • the MRF determines, from the acquired image set, all the images including the information of the object to be determined and the spatio-temporal sequence relationship with the first image and the second image, including : taking a position corresponding to the spatial information of the first image as a starting position, and using a position corresponding to the spatial information of the second image as a termination position, acquiring all the positions from the starting position to the ending position Position information of the image capturing device; generating, according to the relationship between the positions indicated by the position information of all the image capturing devices, the image capturing device corresponding to the starting position as the starting point, and the image capturing device corresponding to the ending position as the end point, generating at least one a device path, wherein each device path includes information of at least one other camera device in addition to the camera device of the start point and the camera device of the end point; for each device path, the time of the first image
  • the time corresponding to the information is the start time
  • the time corresponding to the time information of the second image is the end time, from each of
  • the generating, according to the determined time information and spatial information corresponding to all the images, a prediction path that is the first image as the first node and the second image as the tail node including: a device path, generating, according to the determined time series relationship of the image, a plurality of connected intermediate nodes having a spatio-temporal sequence relationship; generating and the current device path according to the first node, the tail node, and the intermediate node Corresponding image path having a spatiotemporal sequence relationship; determining, from the image path corresponding to each device path, a maximum probability image path with the first image as the head node and the second image as the tail node as the to-be Determine the predicted path of the target.
  • the maximum probability image path with the first image as the first node and the second image as the tail node is determined as the target to be determined from the image path corresponding to each device path.
  • a prediction path comprising: obtaining, for each image path corresponding to each device path, a probability of having information of the same target to be determined between images of two adjacent nodes in the image path; according to each of the image paths a probability of having information of the same target to be determined between images of two adjacent nodes, calculating a probability that the image path is a predicted path of the target to be determined; and each image path is used as a predicted path of the target to be determined The probability of determining the maximum probability image path as the predicted path of the target to be determined.
  • the validity of the prediction path is determined, and based on the determination result, determining whether the target to be determined in the first image and the second image are the same, including: predicting the prediction by using a neural network
  • the path performs validity determination, and based on the determination result, determines whether the target to be determined in the first image and the second image is the same target to be determined.
  • determining, by the neural network, the validity of the prediction path, and determining, according to the determination result, whether the target to be determined in the first image and the second image is the same target to be determined including: Obtaining time information of adjacent images in the prediction path, acquiring time differences of adjacent images; acquiring spatial differences of adjacent images according to spatial information of adjacent images; determining to be determined according to adjacent images Feature information of the target, obtaining feature differences of the target to be determined in the adjacent image; inputting time difference, spatial difference, and feature difference of adjacent images in the predicted path into the long and short time memory network (Long Short-Term Memory, LSTM), obtaining an identification probability of the target to be determined of the predicted path; determining, according to the recognition probability of the target to be determined of the predicted path, whether the target to be determined in the first image and the second image are the same The target is to be determined.
  • LSTM Long Short-Term Memory
  • acquiring feature differences of the target to be determined in the adjacent image according to the feature information of the target to be determined in the adjacent image including: acquiring, by the Siamese-CNN, the to-be-determined in the adjacent image respectively Feature information of the target; acquiring feature differences of the target to be determined in the adjacent image according to the separately acquired feature information.
  • an object recognition apparatus configured to acquire a first image and a second image, where the first image and the second image respectively include a target to be determined; and a generating module configured to be based on the first image and the a second image, a prediction path is generated, the two ends of the prediction path respectively correspond to the first image and the second image; and the first determining module is configured to perform validity determination on the prediction path, based on the determination result Determining whether the target to be determined in the first image and the second image is the same target to be determined.
  • the target to be determined is a vehicle.
  • the device further includes: a second determining module, configured to: according to time information, spatial information, image feature information of the first image, and time information, spatial information of the second image, And determining, by the image feature information, preliminary preliminary probability values of the to-be-determined target respectively included in the first image and the second image; the generating module, comprising: a first generating sub-module configured to perform the preliminary same probability When the value is greater than the preset value, the predicted path is generated based on the first image and the second image.
  • a second determining module configured to: according to time information, spatial information, image feature information of the first image, and time information, spatial information of the second image, And determining, by the image feature information, preliminary preliminary probability values of the to-be-determined target respectively included in the first image and the second image
  • the generating module comprising: a first generating sub-module configured to perform the preliminary same probability When the value is greater than the preset value, the predicted path is generated based on the first image and the second image.
  • the second determining module includes: a first determining submodule configured to: the first image and the second image, and the first image and the second image
  • the difference between the time information and the difference in the spatial information is input to the Siamese-CNN, and preliminary preliminary probability values of the target to be determined in the first image and the second image are obtained.
  • the generating module includes: a second generating submodule configured to: according to feature information of the first image, time information of the first image, spatial information of the first image, The feature information of the second image, the time information of the second image, and the spatial information of the second image are used to generate a predicted path of the target to be determined by using a probability model.
  • the second generation sub-module includes: a first determining unit configured to determine, by the MRF, the information that includes the target to be determined from the acquired image set, and the first image and The second image has all the images of the spatio-temporal sequence relationship; the first generating unit is configured to generate the predicted path of the target to be determined according to the determined time information and spatial information of all the images.
  • the first generating unit includes: a second generating unit, configured to generate, according to the determined time information and spatial information corresponding to all the images, the first image as a first node and the first The two images are a predicted path of the tail node, wherein the predicted path corresponds to at least one intermediate node in addition to the first node and the tail node.
  • the first determining unit is configured to: take a position corresponding to the spatial information of the first image as a starting position, and use a position corresponding to the spatial information of the second image as a ending position to obtain Position information of all the image pickup apparatuses from the start position to the end position; according to the relationship between the positions indicated by the position information of all the image pickup apparatuses, starting from the image pickup apparatus corresponding to the start position, Generating at least one device path with the imaging device corresponding to the termination position as an end point, wherein each device path includes information of at least one other imaging device in addition to the imaging device of the starting point and the imaging device of the end point For each device path, the time corresponding to the time information of the first image is taken as the start time, and the time corresponding to the time information of the second image is the end time, from each other camera device on the current path In the captured image, a map of information including the target to be determined captured by a previous imaging device adjacent to the current imaging device is determined An image having information that sets a time series
  • the second generating unit is configured to generate, according to the determined time series relationship of the image, a plurality of intermediate nodes having a spatio-temporal sequence relationship for each device path;
  • the node, the tail node, and the intermediate node generate an image path having a spatio-temporal sequence relationship corresponding to the current device path; and determining, by the image path corresponding to each device path, that the first image is the first node
  • the maximum probability image path with the second image as the tail node is used as the predicted path of the target to be determined.
  • the second generating unit is further configured to: acquire, for each image path corresponding to each device path, an image of each two adjacent nodes in the image path having the same target to be determined Probability of information; calculating a probability of the image path as a predicted path of the target to be determined according to a probability of having information of the same target to be determined between images of two adjacent nodes in the image path; The image path determines the maximum probability image path as the predicted path of the target to be determined as the probability of the predicted path of the target to be determined.
  • the first determining module includes: a second determining submodule configured to determine validity of the predicted path by using a neural network, and determine the first image and the second according to the determining result. Whether the target to be determined in the image is the same target to be determined.
  • the second determining sub-module includes: a first acquiring unit, configured to acquire a time difference of adjacent images according to time information of adjacent images in the predicted path; The spatial information of the image is obtained, and the spatial difference of the adjacent image is obtained; according to the feature information of the target to be determined in the adjacent image, the feature difference of the target to be determined in the adjacent image is acquired; the second acquiring unit is configured as Inputting the time difference, the spatial difference, and the feature difference of the adjacent images in the obtained prediction path into the LSTM to obtain the recognition probability of the target to be determined of the predicted path; the second determining unit is configured to be according to the predicted path Determining the recognition probability of the target, determining whether the target to be determined in the first image and the second image is the same target to be determined.
  • the first acquiring unit is configured to acquire feature information of the target to be determined in the adjacent image by using Siamese-CNN, and acquire the adjacent image according to the separately obtained feature information. The difference in characteristics of the target to be determined.
  • a computer readable storage medium having stored thereon computer program instructions, wherein the program instructions, when executed by a processor, implement the first aspect of the embodiments of the present invention The steps of the target recognition method.
  • an electronic device comprising: a processor, a memory, a communication component, and a communication bus, wherein the processor, the memory, and the communication component complete each other through the communication bus
  • the memory is for storing at least one executable instruction that causes the processor to perform the steps of the object recognition method according to the first aspect of the embodiments of the present invention.
  • a prediction path that the target to be determined may pass is generated; and the validity of the prediction path is determined to determine the first image and Whether the targets to be determined in the second image are the same.
  • the validity judgment is a possibility of judging whether the current prediction path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher. Thereby, it is possible to perform more accurate detection and recognition on whether the target to be determined in different images is the same target to be determined.
  • FIG. 1 is a schematic flow chart of a target recognition method according to a first embodiment of the present invention
  • FIG. 2 is a schematic flow chart of a target recognition method according to Embodiment 2 of the present invention.
  • FIG. 3 is a schematic flow chart of a target recognition method according to Embodiment 3 of the present invention.
  • FIG. 4 is a block diagram showing the structure of a target recognition apparatus according to Embodiment 4 of the present invention.
  • Figure 5 is a block diagram showing the structure of a target recognition apparatus according to Embodiment 5 of the present invention.
  • FIG. 6 is a structural block diagram of an object recognition apparatus according to Embodiment 6 of the present invention.
  • FIG. 7 is a schematic structural diagram of an electronic device according to Embodiment 7 of the present invention.
  • FIG. 1 is a schematic flow chart of a target recognition method according to a first embodiment of the present invention. As shown in FIG. 1, the target recognition method of this embodiment includes the following steps:
  • step S102 the first image and the second image are acquired.
  • the content to be determined is included in both the first image and the second image from the content included in the image.
  • the first image and the second image may both be captured still images, or video images in a sequence of video frames, and the like.
  • the target to be determined may include a pedestrian, a drone, a vehicle, and the like. It is to be understood that the embodiment is not limited thereto, and any movable object is included in the range of the target to be determined.
  • step S104 a prediction path is generated based on the first image and the second image.
  • the target may be determined based on the feature information of the target to be determined included in the first image and the second image and the spatiotemporal information included in the first image and the second image.
  • the route of travel is predicted, and the reliability of the target identification to be determined is enhanced by the route prediction result.
  • it is necessary to further find a possible travel route of the target to be determined in the image, wherein the target to be determined captured on the travel route The images should all be associated with the first image and the second image in time and space.
  • step S106 the validity of the prediction path is determined, and based on the determination result, it is determined whether the target to be determined in the first image and the second image is the same target to be determined.
  • the validity judgment is a possibility of determining whether a predicted path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined.
  • the result of the validity judgment may specifically be an effective probability, or may be directly “effective”.
  • the object recognition method based on the information included in the first image and the second image, a prediction path through which the target to be determined may pass is generated; and the validity of the prediction path is determined to determine the first image and Whether the targets to be determined in the second image are the same.
  • the validity judgment is a possibility of judging whether the current prediction path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher. Thereby, it is possible to perform more accurate detection and recognition on whether the target to be determined in different images is the same target to be determined.
  • the target recognition method of this embodiment may be performed by any suitable device having image or data processing capability, including but not limited to: camera, terminal, mobile terminal, PC, server, in-vehicle device, entertainment device, advertising device, personal digital Assistants (PDAs), tablets, laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • PDAs personal digital Assistants
  • tablets laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • FIG. 2 a flow chart of a target recognition method according to a second embodiment of the present invention is shown.
  • the object recognition method of the embodiment of the present invention is described by taking the object to be determined as a vehicle as an example, but those skilled in the art should understand that in practical applications, other objects to be determined may refer to the embodiment. Achieve the corresponding target recognition operation.
  • step S202 the first image and the second image are acquired.
  • the first image and the second image each include a target to be determined, and the target to be determined is a vehicle.
  • step S204 according to the feature information of the first image, the time information of the first image, the spatial information of the first image, the feature information of the second image, and the time information of the second image. And the spatial information of the second image, and the predicted path of the target to be determined is generated by a probability model.
  • the travel route of the vehicle is more stable and regular, and the accuracy of judgment and recognition is higher. Therefore, the characteristic information of the vehicle (which can characterize the appearance of the vehicle) and the image can be jointly utilized. Temporal and spatial information to predict the travel route of the vehicle, and to enhance the credibility of the vehicle identification with the route prediction result.
  • the time information of the image is used to indicate the time when the image is captured, and may be regarded as the time when the target (such as a vehicle) to be determined passes the shooting device; the spatial information of the image is used to indicate the position of the captured image, which may be considered as the location of the shooting device.
  • the location may also be considered as the location of the target to be determined when the vehicle is photographed;
  • the feature information of the image is used to indicate a feature of the target to be determined in the image, such as a feature of the vehicle, according to which the vehicle may be determined.
  • Information such as appearance. It can be understood that the information included in the image related to the embodiment may include, but is not limited to, time information of the image, spatial information of the image, and feature information of the image.
  • the probability model can be an MRF.
  • a random field can be thought of as a collection of random variables corresponding to the same sample space. In general, when there is a dependency between these random variables, the random field can be considered to have practical significance.
  • the random field contains two elements, the site and the phase space. When a value of phase space is randomly assigned to each location according to a certain distribution, the whole is called a random field.
  • the MRF is a random field with a Markov nature restriction.
  • the Markov property refers to the distribution of a random variable sequence in chronological order. The distribution characteristics at the N+1th time are independent of the value of the random variable before N.
  • An MRF corresponds to an undirected graph. Each node on the undirected graph corresponds to a random variable, and the edge between the nodes indicates a probability dependency between the random variables corresponding to the node. Therefore, the structure of MRF essentially reflects a priori knowledge, that is, which variables have dependencies to consider and which can be ignored.
  • At least one predicted path of the target to be determined in the first image and the second image may be generated by the MRF, and then the optimal path is determined therefrom as the predicted path of the target to be determined.
  • the MRF may be adopted according to the feature information of the first image, the time information of the first image, the spatial information of the first image, the feature information of the second image, the time information of the second image, and the spatial information of the second image. Generating a predicted path of the target to be determined.
  • all the images including the information of the object to be determined and having a spatio-temporal sequence relationship with the first image and the second image may be determined from the acquired image set by the chain MRF;
  • the time information and the spatial information corresponding to all the images generate a predicted path of the target to be determined.
  • spatio-temporal data refers to data having both time and space dimensions, including information of two dimensions of time and space.
  • spatio-temporal data can be regarded as a time-series set with spatial correlation and a real-time null sequence.
  • the data in the set can be thought of as data with a spatiotemporal sequence relationship.
  • all the images having the spatio-temporal sequence relationship with the first image and the second image have meaning that the spatio-temporal data included in the all images and the spatio-temporal data included in the first image and the spatio-temporal data included in the second image are respectively Time and space are related.
  • the first image is used as the path head node image
  • the second image is the path tail node image
  • the time information and spatial information corresponding to all the images determined by the chain MRF are generated, and the first image is used as the first node.
  • the second image is a predicted path of the tail node, wherein the predicted path corresponds to the at least one intermediate node in addition to the first node and the tail node.
  • the position corresponding to the spatial information of the image is the starting position
  • the position corresponding to the spatial information of the second image is the ending position
  • the position information of all the imaging devices from the starting position to the ending position is acquired; a relationship between the locations indicated by the location information of all the imaging devices, starting from the imaging device corresponding to the initial location, and generating an at least one device path with the imaging device corresponding to the termination location as an end point, wherein each The device path includes information of at least one other imaging device in addition to the imaging device of the starting point and the imaging device of the end point; for each device path, the time corresponding to the time information of the first image is taken as the starting time, The time taken by the time information of the second image is the end time, and the picture taken from each of the other camera devices on the current path , It is determined that the current captured image information imaging apparatus adjacent
  • the determined image may be determined for each device path.
  • the time series relationship generates a plurality of connected intermediate nodes having a spatio-temporal sequence relationship; and according to the first node, the tail node, and the intermediate node, generating an image path having a spatio-temporal sequence relationship corresponding to the current device path;
  • a maximum probability image path with the first image as the first node and the second image as the tail node is determined as the predicted path of the target to be determined.
  • each node on the chain is a camera, and the variable space of the node is taken by the camera.
  • Siamese-CNN can be regarded as the potential energy function of adjacent nodes in MRF.
  • the maximum subsequence and (Max-Sum) algorithm can be used to minimize (optimize) the product value of the potential energy function to obtain the most probable prediction path.
  • the predicted path includes the geographic location of the camera through which the vehicle passes, the time taken, and related information of the captured image.
  • setting p indicates information of the first image (including feature information, time information, and spatial information)
  • q indicates information of the second image (including feature information, time information, and spatial information), which may be from a plurality of pieces by chain MRF
  • One way to determine the optimal path in the predicted path can be achieved by maximizing the following formula (1):
  • P represents the predicted path (ie, the predicted path through which the vehicle is likely to pass);
  • X represents the camera;
  • N represents the number of cameras on a predicted path, from X1 to XN, where x 1 represents the vehicle photographed by X1 Information of the image, and so on, x N represents information of the image of the vehicle captured by the XN,
  • Indicates the potential energy function ie, the output of Siamese-CNN, a probability value between 0 and 1
  • Representing the potential energy function pair between x i and x i+1 , x i and x i+1 are considered to contain information for the same vehicle. If x i and x i+1 do contain information about the same vehicle, then There will be a larger value, otherwise there will be a smaller value.
  • the time constraint described in the formula (2) can be used to make the formula (2) satisfy the formula (3), namely:
  • t is the time, with The optimal selection of the information of the image corresponding to x i and the optimal selection of the information of the image corresponding to x i+1 ;
  • X represents the camera;
  • N represents the number of cameras on a predicted path, from X1 to XN,
  • x N represents information of an image of the vehicle photographed by XN.
  • the information of the image includes time information, spatial information, and feature information of the image.
  • the above formula (1) can be optimized to the following formula (4) to obtain an optimal path, that is, a maximum probability path through which the vehicle can pass.
  • the first image is the prediction path head node A
  • the second image is the prediction path tail node D.
  • the possible driving route of the vehicle includes: route 1: A->B->C -> D; Route 2: A->E->D; Route 3: A->F->G->H->D.
  • step S206 the validity of the prediction path is determined by the neural network, and based on the determination result, whether the target to be determined in the first image and the second image is the same target to be determined is determined.
  • the neural network may be any suitable neural network that can implement feature extraction or target object recognition, including but not limited to a convolutional neural network, an enhanced learning neural network, a generation network in an anti-neural network, and the like.
  • the configuration of the specific structure in the neural network may be appropriately set by a person skilled in the art according to actual needs, such as the number of layers of the convolution layer, the size of the convolution kernel, the number of channels, and the like, which are not limited in the embodiment of the present invention.
  • the neural network can be an LSTM.
  • LSTM is a time recurrent neural network, a variant of a cyclic neural network (RNN) that is better at processing sequence information.
  • RNN cyclic neural network
  • the predicted path of the vehicle may also be considered as a sequence information, which is processed by LSTM to determine the validity of the predicted path.
  • the validity judgment is a possibility of judging whether a predicted path will be the same traveling route of the target to be determined, and the higher the probability, the possibility that the target to be determined in the first image and the second image is the same target to be determined. The higher the sex.
  • the time difference of the adjacent image may be acquired according to the time information of the adjacent image in the prediction path; and the spatial difference of the adjacent image is obtained according to the spatial information of the adjacent image; Obtaining feature information of the target to be determined in the adjacent image, acquiring feature differences of the target to be determined in the adjacent image; inputting time difference, spatial difference, and feature difference of adjacent images in the obtained predicted path into the LSTM, Obtaining a recognition probability of the target to be determined of the predicted path; determining, according to the recognition probability of the target to be determined of the predicted path, whether the target to be determined in the first image and the second image is the same target to be determined.
  • the specific setting of the criterion for determining whether the target is the same target to be determined may be appropriately set by a person skilled in the art according to actual needs, and the embodiment of the present invention does not limit this.
  • the time difference of adjacent images can be obtained by subtracting the time information of the two; the spatial difference of the adjacent images can be obtained by calculating the distance between the two; and the feature difference of the adjacent images can pass the feature vectors of the two Subtraction is obtained.
  • the feature information of the target to be determined in the adjacent image may be respectively acquired by Siamese-CNN by using Siamese-CNN; and the feature information acquired according to the separately acquired feature information is obtained.
  • the Siamese-CNN in this step may be the same as or different from the Siamese-CNN in step S204.
  • the judgment mode adopted is that the LSTM is used for judging, and the input of the LSTM is the time difference (ie, time difference) between the adjacent nodes on the route, the distance difference (ie, the spatial difference), and the difference in appearance thereof. (ie, feature difference), as described above, the difference in appearance can be obtained by directly subtracting the feature vectors output from the two images after input to Siamese-CNN.
  • the output of the LSTM is a probability value by which the validity of the predicted path can be determined to determine whether the vehicles in the two images are indeed the same vehicle.
  • a prediction path that the target to be determined in the image may pass is generated; and the validity of the prediction path is determined to determine the first Whether the target to be determined in an image and the second image is the same.
  • the validity judgment is a possibility of judging whether the current prediction path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher. Thereby, it is possible to perform more accurate detection and recognition on whether the target to be determined in different images is the same target to be determined.
  • the target recognition method of this embodiment may be performed by any suitable device having image or data processing capability, including but not limited to: camera, terminal, mobile terminal, PC, server, in-vehicle device, entertainment device, advertising device, personal digital Assistants (PDAs), tablets, laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • PDAs personal digital Assistants
  • tablets laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • FIG. 3 a flow chart of a target recognition method according to Embodiment 3 of the present invention is shown.
  • the object recognition method of the embodiment of the present invention is described by taking the object to be determined as a vehicle as an example, but those skilled in the art should understand that in practical applications, other objects to be determined may refer to the embodiment. Achieve the corresponding target recognition operation.
  • step S302 the first image and the first image are determined according to time information, spatial information, image feature information of the first image, and time information, spatial information, and image feature information of the second image.
  • the first image and the second image both contain information of the target to be determined.
  • the first image and the second image have a spatio-temporal sequence relationship, and each includes information of a corresponding target to be determined, and based on comprehensive consideration of time information, spatial information, and image feature information of the image, the field
  • the technician can initially determine the preliminary identical probability values of the targets to be determined in the two images by any suitable method.
  • a preliminary identical probability value of the target to be determined respectively included in the first image and the second image may be obtained using Siamese-CNN.
  • Siamese-CNN is a CNN with at least two branches. It can receive multiple inputs at the same time and output the similarity of the multiple inputs (which can be expressed in the form of probability). Taking the double branch as an example, two images can be input simultaneously to the Siamese-CNN through the double branch, and the Siamese-CNN will output the similarity between the two images, or output a judgment result of whether the two images are similar.
  • the Siamese-CNN in this embodiment includes three branches, two of which are used to receive an input image, and one branch is used to receive a difference (time difference) between time information between two images input and a difference in spatial information. (space difference).
  • the similarity (such as appearance similarity) of the target object (the vehicle in this embodiment) in the output image is detected, and the difference of the input time information and the difference of the spatial information are detected.
  • the similarity of the target object in time and space in the output image According to the similarity of the two aspects, the target object in the image, such as the preliminary identical probability value of the vehicle in the present embodiment, can be further determined.
  • the first image and the second image, and the difference of the time information between the first image and the second image and the difference of the spatial information can be input into the Siamese-CNN to obtain the first image.
  • preliminary preliminary probability values of the target to be determined in the second image After obtaining the preliminary same probability value, it is initially determined according to the preliminary same probability value that the first image and the second image have the same target to be determined.
  • comparing the preliminary same probability value with a preset value when the preliminary same probability value is less than or equal to a preset value, determining that the first image and the second image do not have the same target to be determined And determining, when the preliminary same probability value is greater than the preset value, that the first image and the second image have the same target to be determined.
  • the preset value may be appropriately set by a person skilled in the art according to the actual situation, which is not limited by the embodiment of the present invention.
  • Siamese-CNN can effectively judge the degree of similarity of target objects such as vehicles in two images with spatio-temporal information, but not limited to Siamese-CNN.
  • Other methods or neural networks that have similar functions or can achieve the same purpose are also applicable. The solution of the embodiment of the present invention.
  • step S304 when the preliminary same probability value is greater than a preset value, the predicted path is generated based on the first image and the second image.
  • the route to be determined such as the route of the vehicle
  • the vehicle's feature information which can characterize the appearance of the vehicle
  • space-time information can be used in combination to make the route of the vehicle. It is estimated that the credibility of vehicle re-identification will be enhanced by route estimation results.
  • the first image and the second image are images having a spatiotemporal sequence relationship, and based on this, it is necessary to further find a possible travel route of the vehicle in the image, wherein the vehicle photographed on the travel route
  • the images should all have a spatiotemporal sequence relationship with the first image and the second image.
  • the predicted path of the target to be determined is generated using the MRF according to the information of the first image and the information of the second image.
  • the specific implementation process is similar to the step S204 in the foregoing Embodiment 2, and details are not described herein again.
  • step S306 the validity of the prediction path is determined, and based on the determination result, whether the target to be determined in the first image and the second image is the same target to be determined is re-identified.
  • the validity judgment is a possibility of determining whether a predicted path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher.
  • the initially determined result is itself erroneous, that is, the vehicle in the first image and the vehicle in the second image may not be the same vehicle, but are mistaken. Recognized as the same vehicle. If the two are not the same vehicle, the probability that the two have the same driving route within a reasonable time range is very low. Therefore, the prediction path is determined according to the information of the first image and the information of the second image. The validity is also low, whereby re-judgement and recognition of whether the vehicle in the first image and the second image is the same vehicle can be achieved.
  • the validity of the prediction path is determined by using the LSTM, and whether the target to be determined in the first image and the second image is the same target to be determined is re-identified according to the determination result.
  • the specific implementation process is similar to the step S206 in the foregoing Embodiment 2, and details are not described herein again.
  • the object recognition method on the basis of initially determining that the to-be-determined targets respectively included in the first image and the second image are the same, determining a prediction path that the target to be determined may pass; Judging the validity of the prediction path, determining whether the initially determined result is accurate, so as to realize whether the target to be determined in the first image and the second image is re-identified by the same target to be determined.
  • the validity judgment is a possibility of judging whether the current prediction path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher. Thereby, it is possible to perform more accurate re-detection and recognition on whether the target to be determined in different images is the same target to be determined.
  • the target recognition method of this embodiment may be performed by any suitable device having image or data processing capability, including but not limited to: camera, terminal, mobile terminal, PC, server, in-vehicle device, entertainment device, advertising device, personal digital Assistants (PDAs), tablets, laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • PDAs personal digital Assistants
  • tablets laptops, handheld game consoles, smart glasses, smart watches, wearables, virtual display devices or display enhancement devices (such as Google Glass, Oculus Rift, Hololens, Gear VR).
  • FIG. 4 is a block diagram showing the structure of an object recognition apparatus according to Embodiment 4 of the present invention. It can be used to execute the target recognition method flow as described in the first embodiment.
  • the target recognition apparatus includes an acquisition module 401, a generation module 402, and a first determination module 403.
  • the obtaining module 401 is configured to acquire the first image and the second image, where the first image and the second image respectively include a target to be determined;
  • the generating module 402 is configured to generate a prediction path based on the first image and the second image, where two ends of the prediction path respectively correspond to the first image and the second image;
  • the first determining module 403 is configured to perform validity determination on the predicted path, and determine, according to the determination result, whether the target to be determined in the first image and the second image is the same target to be determined.
  • the target recognition apparatus provided in this embodiment generates a prediction path through which the target to be determined may pass based on the information included in the first image and the second image, and determines the first image by determining the validity of the prediction path. Whether the targets to be determined in the second image are the same.
  • the validity judgment is a possibility of judging whether the current prediction path is a travel route of the same target to be determined, and the higher the probability, the target to be determined in the first image and the second image is the same target to be determined. The possibility is also higher. Thereby, it is possible to perform more accurate detection and recognition on whether the target to be determined in different images is the same target to be determined.
  • FIG. 5 is a block diagram showing the structure of a target recognition apparatus according to Embodiment 5 of the present invention. It can be used to perform the target recognition method flow as described in the second embodiment.
  • the target recognition apparatus includes an acquisition module 501, a generation module 502, and a first determination module 503.
  • the acquiring module 501 is configured to acquire the first image and the second image, where the first image and the second image respectively include a target to be determined
  • the generating module 502 is configured to be based on the first image and the a second image, a prediction path is generated, and the two ends of the prediction path respectively correspond to the first image and the second image
  • the first determining module 503 is configured to perform validity determination on the prediction path, based on the determination result Determining whether the target to be determined in the first image and the second image is the same target to be determined.
  • the generating module 502 includes: a second generating submodule 5021, configured to: according to feature information of the first image, time information of the first image, spatial information of the first image And the feature information of the second image, the time information of the second image, and the spatial information of the second image, and the predicted path of the target to be determined is generated by a probability model.
  • the second generation sub-module 5021 includes: a first determining unit 5022, configured to determine information including the target to be determined from the acquired image set by using an MRF, and the first The image and the second image both have all images in a spatiotemporal sequence relationship; the first generating unit 5023 is configured to generate a predicted path of the target to be determined according to the determined time information and spatial information corresponding to all the images.
  • the first generating unit 5023 includes: a second generating unit 5024, configured to generate, according to the determined time information and spatial information of all the images, the first image as a first node and The second image is a predicted path of the tail node, wherein the predicted path corresponds to at least one intermediate node in addition to the first node and the tail node.
  • the first determining unit 5022 is configured to: start a position corresponding to the spatial information of the first image, and use a position corresponding to the spatial information of the second image as a termination position, where Obtaining position information of all the imaging devices from the starting position to the ending position; according to the relationship between the positions indicated by the position information of all the imaging devices, starting from the imaging device corresponding to the starting position Generating at least one device path with the imaging device corresponding to the termination position as an end point, wherein each device path includes at least one other imaging device in addition to the imaging device of the starting point and the imaging device of the end point Information; for each device path, the time corresponding to the time information of the first image is the start time, and the time corresponding to the time information of the second image is the end time, from each other camera on the current path In the image captured by the device, the information of the target to be determined captured by the previous imaging device adjacent to the current imaging device is determined.
  • the image has an image that sets a time series relationship and contains information of the object to be determined
  • the second generating unit 5024 is configured to: for each device path, generate a plurality of connected intermediate nodes having a spatio-temporal sequence relationship according to the determined time series relationship of the image; The first node, the tail node, and the intermediate node generate an image path having a spatio-temporal sequence relationship corresponding to the current device path; and determining, by the image path corresponding to each device path, the first image as a first node And the maximum probability image path with the second image as the tail node is used as the predicted path of the target to be determined.
  • the second generating unit 5024 is further configured to: obtain, for each image path corresponding to each device path, an image of each two adjacent nodes in the image path having the same target to be determined The probability of the information; calculating the probability that the image path is the predicted path of the target to be determined according to the probability of having information of the same target to be determined between the images of every two adjacent nodes in the image path; An image path is used as the probability of the predicted path of the target to be determined to determine a maximum probability image path as the predicted path of the target to be determined.
  • the first determining module 503 includes: a second determining submodule 5031, configured to perform validity determination on the predicted path by using a neural network, and determine the first image and based on the determining result. Whether the target to be determined in the second image is the same target to be determined.
  • the second determining submodule 5031 includes: a first obtaining unit 5032, configured to acquire a time difference of adjacent images according to time information of adjacent images in the predicted path; The spatial information of the adjacent image is obtained, and the spatial difference of the adjacent image is obtained; and the feature difference of the target to be determined in the adjacent image is acquired according to the feature information of the target to be determined in the adjacent image; the second acquiring unit 5033 And configured to input the time difference, the spatial difference, and the feature difference of the adjacent images in the obtained prediction path into the LSTM, to obtain the recognition probability of the target to be determined of the predicted path; and the second determining unit 5034 is configured to Determining, by the recognition probability of the target to be determined of the predicted path, determining whether the target to be determined in the first image and the second image is the same target to be determined.
  • the first acquiring unit 5032 is configured to: respectively acquire feature information of the target to be determined in the adjacent image by using Siamese-CNN; and acquire adjacent images according to the separately acquired feature information. The difference in characteristics of the target to be determined.
  • FIG. 6 is a block diagram showing the structure of an object recognition apparatus according to Embodiment 6 of the present invention. It can be used to perform the target recognition method flow as described in the third embodiment.
  • the target recognition apparatus includes an acquisition module 601, a generation module 603, and a first determination module 604.
  • the obtaining module 601 is configured to acquire the first image and the second image, where the first image and the second image respectively include a target to be determined
  • the generating module 603 is configured to be based on the first image and the a second image, a prediction path is generated, the two ends of the prediction path respectively correspond to the first image and the second image
  • the first determining module 604 is configured to perform validity determination on the prediction path, based on the determination result Determining whether the target to be determined in the first image and the second image are the same.
  • the target to be determined is a vehicle.
  • the apparatus further includes: a second determining module 602 configured to: time information according to the first image, spatial information, image feature information, and time information and spatial information of the second image And determining, by the image feature information, preliminary preliminary probability values of the target to be determined respectively included in the first image and the second image; correspondingly, the generating module 603 includes: a first generating submodule 6031 configured to When the preliminary same probability value is greater than a preset value, the predicted path is generated based on the first image and the second image.
  • the second determining module 602 includes: a first determining submodule 6021 configured to combine the first image and the second image, and the first image and the second The difference in temporal information between the images and the difference in spatial information are input to the Siamese-CNN, and preliminary preliminary probability values of the objects to be determined in the first image and the second image are obtained.
  • Embodiment 7 of the present invention provides an electronic device, which may be, for example, a mobile terminal, a personal computer (PC), a tablet computer, a server, or the like.
  • the electronic device 700 includes one or more processors, communication components Etc., the one or more processors, for example: one or more central processing units (CPUs) 701, and/or one or more image processors (GPUs) 713, etc., may be stored in a read-only memory ( Executable instructions in ROM) 702 or executable instructions loaded from random access memory (RAM) 703 from storage portion 708 perform various appropriate actions and processes.
  • the communication component includes a communication component 712 and/or a communication interface 709.
  • the communication component 712 can include, but is not limited to, a network card, which can include, but is not limited to, an IB (Infiniband) network card
  • the communication interface 709 includes a communication interface of a network interface card such as a LAN card, a modem, etc.
  • the communication interface 709 is via, for example, the Internet.
  • the network performs communication processing.
  • the processor can communicate with read only memory 702 and/or random access memory 703 to execute executable instructions, communicate with communication component 712 over communication bus 704, and communicate with other target devices via communication component 712, thereby completing embodiments of the present invention.
  • Providing an operation corresponding to any one of the target recognition methods for example, acquiring the first image and the second image, wherein the first image and the second image each include a target to be determined; based on the first image and the a second image, a prediction path is generated, the two ends of the prediction path respectively corresponding to the first image and the second image; determining validity of the prediction path, determining the first image based on the determination result Whether the targets to be determined in the second image are the same.
  • RAM 703 various programs and data required for the operation of the device can be stored.
  • the CPU 701 or the GPU 713, the ROM 702, and the RAM 703 are connected to each other through a communication bus 704.
  • ROM 702 is an optional module.
  • the RAM 703 stores executable instructions or writes executable instructions to the ROM 702 at runtime, the executable instructions causing the processor to perform operations corresponding to the above-described communication methods.
  • An input/output (I/O) interface 705 is also coupled to communication bus 704.
  • the communication component 712 can be integrated or can be configured to have multiple sub-modules (e.g., multiple IB network cards) and be on a communication bus link.
  • the following components are connected to the I/O interface 705: an input portion 706 including a keyboard, a mouse, etc.; an output portion 707 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker; a storage portion 708 including a hard disk or the like And a communication interface 709 including a network interface card such as a LAN card, modem, or the like.
  • Driver 710 is also connected to I/O interface 705 as needed.
  • a removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory or the like, is mounted on the drive 710 as needed so that a computer program read therefrom is installed into the storage portion 708 as needed.
  • FIG. 7 is only an optional implementation manner.
  • the number and type of components in FIG. 7 may be selected, deleted, added, or replaced according to actual needs;
  • Different function components can also be implemented in separate settings or integrated settings, such as GPU and CPU detachable settings or GPU can be integrated on the CPU, communication components can be separated, or integrated on the CPU or GPU. and many more.
  • an embodiment of the invention includes a computer program product comprising a computer program tangibly embodied on a machine readable medium, the computer program comprising program code for executing the method illustrated in the flowchart, the program code comprising the corresponding execution
  • the instructions corresponding to the method steps provided by the embodiment of the present invention for example, acquiring the first image and the second image, wherein the first image and the second image respectively include a target to be determined; based on the first image and the a second image, a prediction path is generated, the two ends of the prediction path respectively corresponding to the first image and the second image; determining validity of the prediction path, determining the first image based on the determination result Whether the targets to be determined in the second image are the same.
  • the computer program can be downloaded and installed from the network via a communication component, and/or installed from the removable media 711.
  • the above method according to an embodiment of the present invention may be implemented in hardware, firmware, or implemented as software or computer code that may be stored in a recording medium such as a CD ROM, a RAM, a floppy disk, a hard disk, or a magneto-optical disk, or implemented by
  • the network downloads computer code originally stored in a remote recording medium or non-transitory machine readable medium and stored in a local recording medium so that the methods described herein can be stored using a general purpose computer, a dedicated processor or programmable
  • Such software processing on a recording medium of dedicated hardware such as an ASIC or an FPGA.
  • a computer, processor, microprocessor controller or programmable hardware includes storage components (eg, RAM, ROM, flash memory, etc.) that can store or receive software or computer code, when the software or computer code is The processing methods described herein are implemented when the processor or hardware is accessed and executed. Moreover, when a general purpose computer accesses code for implementing the processing shown herein, the execution of the code converts the general purpose computer into a special purpose computer for performing the processing shown herein.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Neurology (AREA)
  • Algebra (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Image Analysis (AREA)
  • Traffic Control Systems (AREA)

Abstract

一种目标识别方法、装置、存储介质和电子设备。其中，所述目标识别方法包括：获取第一图像及第二图像（S102），所述第一图像及所述第二图像中均包含待确定目标；基于所述第一图像及所述第二图像，生成预测路径（S104），所述预测路径的两端分别对应所述第一图像及所述第二图像；对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标（S106）。

Description

目标识别方法、装置、存储介质和电子设备
相关申请的交叉引用
本申请基于申请号为201710633604.3、申请日为2017年07月28日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此以引入方式并入本申请。
技术领域
本发明实施例涉及人工智能技术领域，尤其涉及一种目标识别方法、装置、存储介质和电子设备。
背景技术
交通工具再识别，如车辆再识别，是计算机视觉和公共安全领域的重要内容，其在交通工具检测与跟踪、行进路线估计以及异常行为检测等诸多方面都有重要应用价值。
大多数交通工具再识别的技术都是基于交通工具的外观信息进行判断，不同于行人再识别，单纯利用交通工具外观信息进行交通工具再识别的难点在于许多交通工具如车辆的外观(如颜色、型号、造型等)十分相似。尤其是在同品牌同款的不同交通工具之间，差异将会更加微小。而依赖交通工具标识信息如车辆的车牌信息以及交通工具内装饰物如车内装饰物这些独特细节信息进行检测和识别，则会因为监控镜头角度不佳、光照条件不佳、镜头模糊等因素使检测和识别的鲁棒性变差，导致检测和识别结果不准确。
发明内容
本发明实施例提供了一种目标识别的技术方案。
根据本发明实施例的第一方面，提供了一种目标识别方法。所述方法包括：获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
在一实施例中，所述待确定目标为交通工具。
在一实施例中，所述基于所述第一图像及所述第二图像，生成预测路径之前，所述方法还包括：根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值；所述基于所述第一图像及所述第二图像，生成预测路径，包括：当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
在一实施例中，根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值，包括：将所述第一图像和所述第二图像，以及，所述第一图像和所述第二图像之间的时间信息的差异和空间信息的差异输入孪生卷积神经网络(Siamese Convolutional Neural Network，Siamese-CNN)，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。
在一实施例中，所述基于所述第一图像及所述第二图像，生成预测路径，包括：根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
在一实施例中，所述通过概率模型生成所述待确定目标的预测路径，包括：通过链状马尔可夫随机场模型(Markov Random Field，MRF)从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像；根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
在一实施例中，所述根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径，包括：根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，其中，所述预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
在一实施例中，所述通过MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像，包括：以所述第一图像的空间信息对应的位置为起始位置，以所述第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；针对每一个设备路径，以所述第一图像 的时间信息对应的时间为起始时间，以所述第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
在一实施例中，所述根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，包括：针对每一个设备路径，根据确定出的所述图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，所述从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径，包括：针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否相同，包括：通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标，包括：根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入长短时记忆网络(Long Short-Term Memory，LSTM)，获得所述预测路径的待确定目标的识别概率；根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异，包括：通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；根据分别获取的所述特征信息，获取相邻的图像中的待确定目标的特征差异。
根据本发明实施例的第二方面，提供了一种目标识别装置。所述装置包括：获取模块，配置为获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；生成模块，配置为基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；第一确定模块，配置为对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
在一实施例中，所述待确定目标为交通工具。
在一实施例中，所述装置还包括：第二确定模块，配置为根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值；所述生成模块，包括：第一生成子模块，配置为当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
在一实施例中，所述第二确定模块，包括：第一确定子模块，配置为将所述第一图像和所述第二图像，以及，所述第一图像和所述第二图像之间的时间信息的差异和空间信息的差异输入Siamese-CNN，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。
在一实施例中，所述生成模块，包括：第二生成子模块，配置为根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
在一实施例中，所述第二生成子模块，包括：第一确定单元，配置为通过MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像；第一生成单元，配置为根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
在一实施例中，所述第一生成单元，包括：第二生成单元，配置为根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，其中，所述预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
在一实施例中，所述第一确定单元，配置为：以所述第一图像的空间信息对应的位置为起始位置，以所述第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；针对每 一个设备路径，以所述第一图像的时间信息对应的时间为起始时间，以所述第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
在一实施例中，所述第二生成单元，配置为：针对每一个设备路径，根据确定出的所述图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，所述第二生成单元，还配置为：针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，所述第一确定模块，包括：第二确定子模块，配置为通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，所述第二确定子模块，包括：第一获取单元，配置为根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；第二获取单元，配置为将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入LSTM，获得所述预测路径的待确定目标的识别概率；第二确定单元，配置为根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，所述第一获取单元，配置为：通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；根据分别获取的所述特征信息，获取相邻的图像中的待确定目标的特征差异。
根据本发明实施例的第三方面，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，其中，所述程序指令被处理器执行时实现本发明实施例的第一方面所述的目标识别方法的步骤。
根据本发明实施例的第四方面，提供了一种电子设备，包括：处理器、存储器、通信元件和通信总线，所述处理器、所述存储器和所述通信元件通过所述通信总线完成相互间的通信；所述存储器用于存放至少一可执行 指令，所述可执行指令使所述处理器执行如本发明实施例的第一方面所述的目标识别方法的步骤。
根据本发明实施例提供的技术方案，基于第一图像和第二图像所包含的信息，生成待确定目标可能通过的预测路径；并通过对该预测路径的有效性判断，以确定第一图像及第二图像中的待确定目标是否相同。其中，有效性判断是对当前预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。由此，可以对不同图像中的待确定目标是否为同一个待确定目标进行较为精准的检测和识别。
附图说明
图1是根据本发明实施例一的一种目标识别方法的流程示意图；
图2是根据本发明实施例二的一种目标识别方法的流程示意图；
图3是根据本发明实施例三的一种目标识别方法的流程示意图；
图4是根据本发明实施例四的一种目标识别装置的结构框图；
图5是根据本发明实施例五的一种目标识别装置的结构框图；
图6是根据本发明实施例六的一种目标识别装置的结构框图；
图7是根据本发明实施例七的一种电子设备的结构示意图。
具体实施方式
下面结合附图(若干附图中相同的标号表示相同的元素)和实施例，对本发明实施例的具体实施方式作进一步详细说明。以下实施例用于说明本发明，但不用来限制本发明的范围。
本领域技术人员可以理解，本发明实施例中的“第一”、“第二”等术语仅用于区别不同步骤、设备或模块等，既不代表任何特定技术含义，也不表示它们之间的必然逻辑顺序。
实施例一
图1是根据本发明实施例一的一种目标识别方法的流程示意图。如图1所示，本实施例的目标识别方法包括以下步骤：
在步骤S102中，获取第一图像及第二图像。
在具体的实施方式中，从图像包含的内容来讲，所述第一图像和所述第二图像中均包含待确定目标。从图像的类别来讲，所述第一图像和所述第二图像均可为拍摄的静态图像，或者为视频帧序列中的视频图像等。具体地，所述待确定目标可包括行人、无人机以及交通工具等。可以理解的是，本实施例不限于此，任何可移动的物体均被包括在待确定目标的范围内。
在步骤S104中，基于所述第一图像及所述第二图像，生成预测路径。
其中，所述预测路径的两端分别对应所述第一图像及所述第二图像。在本发明实施例中，可基于所述第一图像和所述第二图像中包含的待确定目标的特征信息以及所述第一图像和所述第二图像中包含的时空信息来对待确定目标的行进路线进行预测，以路线预测结果来加强待确定目标识别的可信度。具体地，以所述第一图像和所述第二图像所包含的信息为基础，需要进一步找出图像中的待确定目标的可能的行进路线，其中，该行进路线上拍摄的待确定目标的图像均应与所述第一图像和所述第二图像在时空上相关。
在步骤S106中，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
其中，有效性判断是对一条预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高，也即是所述第一图像中的待确定目标与所述第二图像中的待确定目标相同的可能性也越高。在具体的实施方式中，所述有效性判断的结果具体可以是有效概率，也可以直接是“是否有效”。
根据本实施例提供的目标识别方法，基于第一图像和第二图像所包含的信息，生成待确定目标可能通过的预测路径；并通过对该预测路径的有效性判断，以确定第一图像及第二图像中的待确定目标是否相同。其中，有效性判断是对当前预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。由此，可以对不同图像中的待确定目标是否为同一个待确定目标进行较为精准的检测和识别。
本实施例的目标识别方法可以由任意适当的具有图像或数据处理能力的设备执行，包括但不限于：摄像头、终端、移动终端、PC机、服务器、车载设备、娱乐设备、广告设备、个人数码助理(PDA)、平板电脑、笔记本电脑、掌上游戏机、智能眼镜、智能手表、可穿戴设备、虚拟显示设备或显示增强设备(如Google Glass、Oculus Rift、Hololens、Gear VR)等。
实施例二
参照图2，示出了根据本发明实施例二的一种目标识别方法的流程示意图。
在本实施例中，以待确定目标为交通工具为例，对本发明实施例的目标识别方法进行说明，但本领域技术人员应当明了，在实际应用中，其它的待确定目标可参照本实施例实现相应的目标识别操作。
本实施例的目标识别方法包括以下步骤：
在步骤S202中，获取第一图像及第二图像。
在具体的实施方式中，所述第一图像和所述第二图像中均包含待确定目标，且所述待确定目标为交通工具。
在步骤S204中，根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
相比较于行人的行进路线，交通工具的行进路线更加稳定、更加有规律，判断和识别的准确率更高，因此，可以联合利用交通工具的特征信息(可以表征交通工具的外观)和图像的时空信息来对交通工具的行进路线进行预测，以路线预测结果来加强交通工具识别的可信度。
其中，图像的时间信息用于指示拍摄图像的时间，可以认为是待确定目标(如交通工具)经过拍摄设备的时间；图像的空间信息用于指示拍摄图像的位置，可以认为是拍摄设备所在的位置，也可以认为是待确定目标如交通工具被拍摄时所处的位置；图像的特征信息用于指示图像中的待确定目标的特征，如交通工具的特征，根据该特征可以确定交通工具的外观等信息。可以理解的是，本实施例涉及的图像所包含的信息可包括但不限于：图像的时间信息、图像的空间信息和图像的特征信息。
在具体的实施方式中，所述概率模型可为MRF。
随机场可以看成是一组对应于同一个样本空间的随机变量的集合。一般来说，当这些随机变量之间有依赖关系时，该随机场可以被认为具有实际意义。随机场包含两个要素，即位置(site)和相空间(phase space)，当给每一个位置中按照某种分布随机赋予相空间的一个值之后，其全体就叫做随机场。
MRF是加了Markov性质限制的随机场。Markov性质是指一个随机变量序列按时间先后关系依次排开的时候，第N+1时刻的分布特性，与N时刻以前的随机变量的取值无关。一个MRF对应一个无向图，这个无向图上的每一个节点对应一个随机变量，节点之间的边表示节点对应的随机变量之间有概率依赖关系。因此，MRF的结构本质上反应了先验知识，即，哪些变量之间有依赖关系需要考虑，而哪些可以忽略。
本实施例中，可以通过MRF生成第一图像和第二图像中的待确定目标的至少一条预测路径，然后，从中确定出最优路径作为所述待确定目标的预测路径。具体地，可以根据第一图像的特征信息、第一图像的时间信息、第一图像的空间信息、第二图像的特征信息、第二图像的时间信息、和第二图像的空间信息，通过MRF生成所述待确定目标的预测路径。在一实施例中，可以通过链状MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与第一图像和第二图像均具有时空序列关系的所有图像；根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
其中，时空数据是指同时具有时间和空间维度的数据，其包括时间和空间二个维度的信息。地理学中，由于连续的时空数据都是经过离散化抽 样提取并存储的，因而能够将时空数据看作是空间上有相关关系的时间序列集合，即时空序列。该集合中的数据可以被认为是具有时空序列关系的数据。具体地，与第一图像和第二图像均具有时空序列关系的所有图像的含义为该所有图像中包含的时空数据分别与第一图像中包含的时空数据和第二图像中包含的时空数据在时间和空间上相关。
通常情况下，可以以第一图像为路径首节点图像，以第二图像为路径尾节点图像，根据链状MRF确定出的所有图像对应的时间信息和空间信息，生成以第一图像为首节点且以第二图像为尾节点的一条预测路径，其中，该预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
其中，在通过链状MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像时，可以以第一图像的空间信息对应的位置为起始位置，以第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；针对每一个设备路径，以第一图像的时间信息对应的时间为起始时间，以第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
进而，在根据确定出的所有图像对应的时间信息和空间信息，生成以第一图像为首节点且以第二图像为尾节点的一条预测路径时，可以针对每一个设备路径，根据确定出的图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；从每一个设备路径对应的图像路径中，确定出以第一图像为首节点且以第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
其中，在从每一个设备路径对应的图像路径中，确定出以第一图像为首节点且以第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径时，可以针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
以本实施例中的待确定目标为交通工具为例，假定交通工具在路网中 的行进路线为一个链状MRF，链上的每一个节点为一个摄像机，节点的变量空间为该摄像机所拍摄的图像、图像的拍摄时间和地点所组成的三元组。任给一对需要识别是否为同一辆交通工具的图像，以及它们之间的可能的监控摄像机(可能的监控摄像机为先验信息，可以通过任意适当方式得到，如通过在数据训练集上进行统计得到)，把相邻摄像机之间每一对图像以及它们之间的时空差异输入到Siamese-CNN中，计算出每一对路网中相邻监控摄像机所拍的图像中的交通工具所属为同一辆交通工具的概率。其中，Siamese-CNN可以视作为MRF中相邻节点的势能函数，可以通过最大子序列和(Max-Sum)算法，最小化(优化)势能函数的乘积数值，以得到一条可能性最高的预测路径，该预测路径包括该交通工具所经过的摄像机的地理位置、被拍摄的时间以及被拍摄的图像的相关信息。
例如，设定p表示第一图像的信息(包括特征信息、时间信息和空间信息)，q表示第二图像的信息(包括特征信息、时间信息和空间信息)，通过链状MRF从多条可能的预测路径中确定最优路径的一种方式可以采用最大化下述公式(1)的方式实现：
Figure PCTCN2018097374-appb-000001
其中，P表示预测路径(即交通工具有可能经过的预测路径)；X表示摄像机；N表示一条预测路径上的摄像机的数量，从X1到XN，x 1表示X1拍摄到的所述交通工具的图像的信息，以此类推，x N表示XN拍摄到的所述交通工具的图像的信息，
Figure PCTCN2018097374-appb-000002
表示势能函数(即Siamese-CNN的输出，为0～1之间的概率值)，
Figure PCTCN2018097374-appb-000003
表示x i和x i+1之间的势能函数对，x i和x i+1被认为包含有同一辆交通工具的信息。如果x i和x i+1确实包含有同一辆交通工具的信息，则
Figure PCTCN2018097374-appb-000004
将有一个较大值，否则，会有一个较小值。
在最大化上述公式(1)时，可以使用公式(2)中所述的时间约束，使公式(2)满足公式(3)，即：
Figure PCTCN2018097374-appb-000005
Figure PCTCN2018097374-appb-000006
其中，t表示时间，
Figure PCTCN2018097374-appb-000007
和
Figure PCTCN2018097374-appb-000008
分别表示x i对应的图像的信息的最优选择和x i+1对应的图像的信息的最优选择；X表示摄像机；N表示一条预测路径上的摄像机的数量，从X1到XN，x 1表示X1拍摄到的所述交通工具的图像的信息，以此类推，x N表示XN拍摄到的所述交通工具的图像的信息。
上述公式(1)、(2)和(3)中，图像的信息均包括图像的时间信息、空间信息和特征信息。
基于上述公式(1)、(2)和(3)，可以将上述公式(1)最优化为下述公式(4)，以获得最优路径，也即，交通工具可能通过的最大概率路径。
Figure PCTCN2018097374-appb-000009
通过上述过程，可以确定出所述交通工具最有可能经过的一条预测路径。
例如，以第一图像为预测路径首节点A，以第二图像为预测路径尾节点D，根据摄像设备之间的位置关系，车辆可能的行驶路线包括：路线1：A->B->C->D；路线2：A->E->D；路线3：A->F->G->H->D。经过上述公式(4)的计算，确定路线1的概率为85％，路线2的概率为95％，而路线3的概率为70％，则可将路线2确定为交通工具的预测路径。
需要说明的是，上述过程以链状MRF为例，但在实际应用中，本领域技术人员也可以使用其它适当方式实现所述待确定目标的预测路径的生成。例如，根据深度神经网络检测第一图像和第二图像的背景信息，以生成所述待确定目标的预测路径。
在步骤S206中，通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
其中，所述神经网络可以是任意适当的可实现特征提取或目标对象识别的神经网络，包括但不限于卷积神经网络、增强学习神经网络、对抗神经网络中的生成网络等等。神经网络中具体结构的设置可以由本领域技术人员根据实际需求适当设定，如卷积层的层数、卷积核的大小、通道数等等，本发明实施例对此不作限制。
在具体的实施方式中，所述神经网络可为LSTM。LSTM是一种时间递归神经网络，是一种循环神经网络(RNN)的变种，较为擅长于处理序列信息。本发明实施例中，交通工具的预测路径也可以认为是一个序列信息，采用LSTM对其进行处理，以确定预测路径的有效性。
有效性判断是对一条预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。
本实施例中，可以根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入LSTM，获得所述预测路径的待确定目标的识别概率；根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。其中，对待 确定目标是否为同一个待确定目标的判断标准的具体设定可以由本领域技术人员根据实际需要适当设置，本发明实施例对此不作限制。
其中，相邻图像的时间差异可以通过两者的时间信息相减得到；相邻图像的空间差异可以通过计算两者之间的距离得到；而相邻图像的特征差异可以通过两者的特征向量相减得到。在一种可行方式中，在获得相邻图像的特征差异时，可以利用Siamese-CNN，通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；根据分别获取的特征信息，获取相邻的图像中的待确定目标的特征差异。其中，本步骤中的Siamese-CNN可以与步骤S204中的Siamese-CNN相同也可以不同。
本实施例中，在通过MRF得到任意两张交通工具图像之间的行进路线后，需要对该行进路线是否有效进行判断，即进行有效性判断。其中，有效指该行进路线是一条同样的交通工具会走过的路线，反之则为无效路线。本实施例中，采用的判断方式是使用LSTM来进行判断，该LSTM的输入为路线上相邻节点之间的时间差(即时间差异)、距离差(即空间差异)，还有它们的外观差异(即特征差异)，如前所述，其中的外观差异可用两张图像输入到Siamese-CNN后输出的特征向量直接进行相减得到。该LSTM的输出为一个概率值，通过该概率值可以对预测路径是否有效进行判断，进而判断两张图像中的交通工具是否确实为同一辆交通工具。
可见，通过本实施例，基于第一图像和第二图像所包含的时空信息和特征信息，生成图像中待确定目标可能通过的预测路径；并通过对该预测路径的有效性判断，以确定第一图像及第二图像中的待确定目标是否相同。其中，有效性判断是对当前预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。由此，可以对不同图像中的待确定目标是否为同一个待确定目标进行较为精准的检测和识别。
本实施例的目标识别方法可以由任意适当的具有图像或数据处理能力的设备执行，包括但不限于：摄像头、终端、移动终端、PC机、服务器、车载设备、娱乐设备、广告设备、个人数码助理(PDA)、平板电脑、笔记本电脑、掌上游戏机、智能眼镜、智能手表、可穿戴设备、虚拟显示设备或显示增强设备(如Google Glass、Oculus Rift、Hololens、Gear VR)等。
实施例三
参照图3，示出了根据本发明实施例三的一种目标识别方法的流程示意图。在本实施例中，以待确定目标为交通工具为例，对本发明实施例的目标识别方法进行说明，但本领域技术人员应当明了，在实际应用中，其它的待确定目标可参照本实施例实现相应的目标识别操作。
本实施例的目标识别方法包括以下步骤：
在步骤S302中，根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定 所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值。
其中，第一图像和第二图像均包含有待确定目标的信息。
本发明实施例中，第一图像和第二图像具有时空序列关系，且均包含有相应的待确定目标的信息，在综合考虑图像的时间信息、空间信息、图像特征信息的基础上，本领域技术人员可以采用任意适当的方法初步确定这两张图像中的待确定目标的初步相同概率值。
在一种可行方案中，可以使用Siamese-CNN获得所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值。
Siamese-CNN是一种至少具有两个分支的CNN，可以同时接收多个输入，输出该多个输入的相似度(可以表现为概率的形式)。以双分支为例，可以通过双分支同时向Siamese-CNN输入两张图像，Siamese-CNN将输出这两张图像之间的相似度，或者，输出两张图像是否相似的判断结果。本实施例中的Siamese-CNN包括三个分支，其中两个分支用于接收输入的图像，一个分支用于接收输入的两个图像之间的时间信息的差异(时间差异)和空间信息的差异(空间差异)。通过对输入的图像进行检测输出图像中目标对象(本实施例中为交通工具)在特征方面的相似度(如外观相似度)，以及，对输入的时间信息的差异和空间信息的差异进行检测输出图像中目标对象在时空方面的相似度。根据这两方面的相似度，可以进一步确定图像中的目标对象，如本实施例中的交通工具的初步相同概率值。
可见，在本实施例中，可以将第一图像和第二图像，以及，第一图像和第二图像之间的时间信息的差异和空间信息的差异输入Siamese-CNN，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。在获得初步相同概率值之后，根据所述初步相同概率值初步确定所述第一图像和第二图像中具有相同的待确定目标。具体地，将所述初步相同概率值与预设值进行比较，当所述初步相同概率值小于或等于预设值时，确定所述第一图像和第二图像中不具有相同的待确定目标，当所述初步相同概率值大于预设值时，初步确定所述第一图像和第二图像中具有相同的待确定目标。其中，所述预设值可以由本领域技术人员根据实际情况适当设置，本发明实施例对此不作限制。
Siamese-CNN可以对具有时空信息的两张图像中的目标对象如交通工具的相似程度进行有效判断，但不限于Siamese-CNN，其它具有类似功能或者可以实现相同目的的方式或神经网络也同样适用于本发明实施例的方案。
在步骤S304中，当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
相比较于行人的行进路线，待确定目标如交通工具的行进路线更加有规律性，因此，可以联合利用交通工具的特征信息(可以表征交通工具的 外观)和时空信息来对交通工具的路线进行估计，以路线估计结果来加强交通工具再识别的可信度。
如前所述，第一图像和第二图像为具有时空序列关系的图像，以此为基础，需要进一步找出图像中的交通工具的可能的行进路线，其中，该行进路线上拍摄的交通工具的图像均应与第一图像和第二图像具有时空序列关系。
在具体的实施方式中，根据第一图像的信息和第二图像的信息，使用MRF生成所述待确定目标的预测路径。具体的实现过程与上述实施二中的步骤S204类似，在此不再赘述。
在步骤S306中，对所述预测路径进行有效性判断，基于判断结果，对所述第一图像和所述第二图像中的待确定目标是否为同一个待确定目标进行再识别。
其中，有效性判断是对一条预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。
例如，在某些情况下，有可能初步确定的结果本身就是错误的，也即，第一图像中的交通工具和第二图像中的交通工具可能并不为同一辆交通工具，但却被误识别为同一辆交通工具。若二者不为同一辆交通工具，则二者在可能的合理时间范围内具有相同的行驶路线的概率很低，因此，根据第一图像的信息和第二图像的信息确定有的预测路径的有效性也较低，由此可以实现对第一图像和第二图像中的交通工具是否为同一辆交通工具的再判断和识别。
在具体的实施方式中，通过LSTM，对所述预测路径进行有效性判断，根据判断结果对所述第一图像和第二图像中的待确定目标是否为同一个待确定目标进行再识别。具体的实现过程与上述实施二中的步骤S206类似，在此不再赘述。
根据本实施例提供的目标识别方法，在初步确定所述第一图像及所述第二图像中分别包含的待确定目标相同的基础上，确定该待确定目标可能通过的预测路径；进而，通过对该预测路径的有效性判断，确定初步确定的结果是否准确，以实现对第一图像和第二图像中的待确定目标是否为同一个待确定目标的再识别。其中，有效性判断是对当前预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。由此，可以对不同图像中的待确定目标是否为同一个待确定目标进行较为精准的再检测和识别。
本实施例的目标识别方法可以由任意适当的具有图像或数据处理能力的设备执行，包括但不限于：摄像头、终端、移动终端、PC机、服务器、车载设备、娱乐设备、广告设备、个人数码助理(PDA)、平板电脑、笔记 本电脑、掌上游戏机、智能眼镜、智能手表、可穿戴设备、虚拟显示设备或显示增强设备(如Google Glass、Oculus Rift、Hololens、Gear VR)等。
实施例四
基于相同的技术构思，图4是示出根据本发明实施例四的一种目标识别装置的结构示意图。可用以执行如实施例一所述的目标识别方法流程。
参照图4，该目标识别装置包括获取模块401、生成模块402和第一确定模块403。
获取模块401，配置为获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；
生成模块402，配置为基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；
第一确定模块403，配置为对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
通过本实施例提供的目标识别装置，基于第一图像和第二图像所包含的信息，生成待确定目标可能通过的预测路径；并通过对该预测路径的有效性判断，以确定第一图像及第二图像中的待确定目标是否相同。其中，有效性判断是对当前预测路径是否会为同一个待确定目标的行进路线的可能性判断，其可能性越高则第一图像和第二图像中的待确定目标为同一个待确定目标的可能性也越高。由此，可以对不同图像中的待确定目标是否为同一个待确定目标进行较为精准的检测和识别。
实施例五
基于相同的技术构思，图5是示出根据本发明实施例五的一种目标识别装置的结构示意图。可用以执行如实施例二所述的目标识别方法流程。
参照图5，该目标识别装置包括获取模块501、生成模块502和第一确定模块503。其中，获取模块501，配置为获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；生成模块502，配置为基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；第一确定模块503，配置为对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
在一实施例中，所述生成模块502，包括：第二生成子模块5021，配置为根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
在一实施例中，所述第二生成子模块5021，包括：第一确定单元5022，配置为通过MRF从获取的图像集中确定出包含有所述待确定目标的信息、 且与所述第一图像和所述第二图像均具有时空序列关系的所有图像；第一生成单元5023，配置为根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
在一实施例中，所述第一生成单元5023，包括：第二生成单元5024，配置为根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，其中，所述预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
在一实施例中，所述第一确定单元5022，配置为：以所述第一图像的空间信息对应的位置为起始位置，以所述第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；针对每一个设备路径，以所述第一图像的时间信息对应的时间为起始时间，以所述第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
在一实施例中，所述第二生成单元5024，配置为：针对每一个设备路径，根据确定出的所述图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，所述第二生成单元5024，还配置为：针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
在一实施例中，所述第一确定模块503，包括：第二确定子模块5031，配置为通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，所述第二确定子模块5031，包括：第一获取单元5032，配置为根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根 据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；第二获取单元5033，配置为将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入LSTM，获得所述预测路径的待确定目标的识别概率；第二确定单元5034，配置为根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
在一实施例中，所述第一获取单元5032，配置为：通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；根据分别获取的所述特征信息，获取相邻的图像中的待确定目标的特征差异。
需要说明的是，对于本发明实施例提供的目标识别装置还涉及的具体细节已在本发明实施例提供的目标识别方法中作了详细的说明，在此不在赘述。
实施例六
基于相同的技术构思，图6是示出根据本发明实施例六的一种目标识别装置的结构示意图。可用以执行如实施例三所述的目标识别方法流程。
参照图6，该目标识别装置包括获取模块601、生成模块603和第一确定模块604。其中，获取模块601，配置为获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；生成模块603，配置为基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；第一确定模块604，配置为对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否相同。
在一实施例中，所述待确定目标为交通工具。
在一实施例中，所述装置还包括：第二确定模块602，配置为根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值；相应地，所述生成模块603，包括：第一生成子模块6031，配置为当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
在一实施例中，所述第二确定模块602，包括：第一确定子模块6021，配置为将所述第一图像和所述第二图像，以及，所述第一图像和所述第二图像之间的时间信息的差异和空间信息的差异输入Siamese-CNN，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。
需要说明的是，对于本发明实施例提供的目标识别装置还涉及的具体细节已在本发明实施例提供的目标识别方法中作了详细的说明，在此不在赘述。
实施例七
本发明实施例七提供了一种电子设备，例如可以是移动终端、个人计 算机(PC)、平板电脑、服务器等。下面参考图7，其示出了适于用来实现本发明实施例的终端设备或服务器的电子设备700的结构示意图：如图7所示，电子设备700包括一个或多个处理器、通信元件等，所述一个或多个处理器例如：一个或多个中央处理单元(CPU)701，和/或一个或多个图像处理器(GPU)713等，处理器可以根据存储在只读存储器(ROM)702中的可执行指令或者从存储部分708加载到随机访问存储器(RAM)703中的可执行指令而执行各种适当的动作和处理。通信元件包括通信组件712和/或通信接口709。其中，通信组件712可包括但不限于网卡，所述网卡可包括但不限于IB(Infiniband)网卡，通信接口709包括诸如LAN卡、调制解调器等的网络接口卡的通信接口，通信接口709经由诸如因特网的网络执行通信处理。
处理器可与只读存储器702和/或随机访问存储器703中通信以执行可执行指令，通过通信总线704与通信组件712相连、并经通信组件712与其他目标设备通信，从而完成本发明实施例提供的任一项目标识别方法对应的操作，例如，获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否相同。
此外，在RAM 703中，还可存储有装置操作所需的各种程序和数据。CPU701或GPU713、ROM702以及RAM703通过通信总线704彼此相连。在有RAM703的情况下，ROM702为可选模块。RAM703存储可执行指令，或在运行时向ROM702中写入可执行指令，可执行指令使处理器执行上述通信方法对应的操作。输入/输出(I/O)接口705也连接至通信总线704。通信组件712可以集成设置，也可以设置为具有多个子模块(例如多个IB网卡)，并在通信总线链接上。
以下部件连接至I/O接口705：包括键盘、鼠标等的输入部分706；包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分707；包括硬盘等的存储部分708；以及包括诸如LAN卡、调制解调器等的网络接口卡的通信接口709。驱动器710也根据需要连接至I/O接口705。可拆卸介质711，诸如磁盘、光盘、磁光盘、半导体存储器等等，根据需要安装在驱动器710上，以便于从其上读出的计算机程序根据需要被安装入存储部分708。
需要说明的，如图7所示的架构仅为一种可选实现方式，在具体实践过程中，可根据实际需要对上述图7的部件数量和类型进行选择、删减、增加或替换；在不同功能部件设置上，也可采用分离设置或集成设置等实现方式，例如GPU和CPU可分离设置或者可将GPU集成在CPU上，通信元件可分离设置，也可集成设置在CPU或GPU上，等等。这些可替换的 实施方式均落入本发明的保护范围。
特别地，根据本发明实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本发明实施例包括一种计算机程序产品，其包括有形地包含在机器可读介质上的计算机程序，计算机程序包含用于执行流程图所示的方法的程序代码，程序代码可包括对应执行本发明实施例提供的方法步骤对应的指令，例如，获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否相同。在这样的实施例中，该计算机程序可以通过通信元件从网络上被下载和安装，和/或从可拆卸介质711被安装。在该计算机程序被处理器执行时，执行本发明实施例的方法中限定的上述功能。
需要指出，根据实施的需要，可将本发明实施例中描述的各个部件/步骤拆分为更多部件/步骤，也可将两个或多个部件/步骤或者部件/步骤的部分操作组合成新的部件/步骤，以实现本发明实施例的目的。
上述根据本发明实施例的方法可在硬件、固件中实现，或者被实现为可存储在记录介质(诸如CD ROM、RAM、软盘、硬盘或磁光盘)中的软件或计算机代码，或者被实现通过网络下载的原始存储在远程记录介质或非暂时机器可读介质中并将被存储在本地记录介质中的计算机代码，从而在此描述的方法可被存储在使用通用计算机、专用处理器或者可编程或专用硬件(诸如ASIC或FPGA)的记录介质上的这样的软件处理。可以理解，计算机、处理器、微处理器控制器或可编程硬件包括可存储或接收软件或计算机代码的存储组件(例如，RAM、ROM、闪存等)，当所述软件或计算机代码被计算机、处理器或硬件访问且执行时，实现在此描述的处理方法。此外，当通用计算机访问用于实现在此示出的处理的代码时，代码的执行将通用计算机转换为用于执行在此示出的处理的专用计算机。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及方法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明实施例的范围。
以上实施方式仅用于说明本发明实施例，而并非对本发明实施例的限制，有关技术领域的普通技术人员，在不脱离本发明实施例的精神和范围的情况下，还可以做出各种变化和变型，因此所有等同的技术方案也属于本发明实施例的范畴，本发明实施例的专利保护范围应由权利要求限定。

Claims (28)

  1. 一种目标识别方法，包括：
    获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；
    基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；
    对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
  2. 根据权利要求1所述的方法，其中，所述待确定目标为交通工具。
  3. 根据权利要求1所述的方法，其中，所述基于所述第一图像及所述第二图像，生成预测路径之前，所述方法还包括：
    根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值；
    所述基于所述第一图像及所述第二图像，生成预测路径，包括：
    当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
  4. 根据权利要求3所述的方法，其中，根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息，确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值，包括：
    将所述第一图像和所述第二图像，以及，所述第一图像和所述第二图像之间的时间信息的差异和空间信息的差异输入孪生卷积神经网络Siamese-CNN，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。
  5. 根据权利要求1-4任一项所述的方法，其中，所述基于所述第一图像及所述第二图像，生成预测路径，包括：
    根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
  6. 根据权利要求5所述的方法，其中，所述通过概率模型生成所述待确定目标的预测路径，包括：
    通过链状马尔可夫随机场模型MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像；
    根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
  7. 根据权利要求6所述的方法，其中，所述根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径，包括：
    根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，其中，所述预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
  8. 根据权利要求7所述的方法，其中，所述通过链状马尔可夫随机场模型MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像，包括：
    以所述第一图像的空间信息对应的位置为起始位置，以所述第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；
    根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；
    针对每一个设备路径，以所述第一图像的时间信息对应的时间为起始时间，以所述第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
  9. 根据权利要求8所述的方法，其中，所述根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，包括：
    针对每一个设备路径，根据确定出的所述图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；
    从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
  10. 根据权利要求9所述的方法，其中，所述从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径，包括：
    针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；
    根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目 标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；
    根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
  11. 根据权利要求1-10任一项所述的方法，其中，所述对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标，包括：
    通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
  12. 根据权利要求11所述的方法，其中，所述通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标，包括：
    根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；
    将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入长短时记忆网络LSTM，获得所述预测路径的待确定目标的识别概率；
    根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
  13. 根据权利要求12所述的方法，其中，根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异，包括：
    通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；
    根据分别获取的所述特征信息，获取相邻的图像中的待确定目标的特征差异。
  14. 一种目标识别装置，包括：
    获取模块，配置为获取第一图像及第二图像，所述第一图像及所述第二图像中均包含待确定目标；
    生成模块，配置为基于所述第一图像及所述第二图像，生成预测路径，所述预测路径的两端分别对应所述第一图像及所述第二图像；
    第一确定模块，配置为对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像及所述第二图像中的待确定目标是否为同一待确定目标。
  15. 根据权利要求14所述的装置，其中，所述待确定目标为交通工具。
  16. 根据权利要求14所述的装置，其中，所述装置还包括：
    第二确定模块，配置为根据所述第一图像的时间信息、空间信息、图像特征信息，以及，所述第二图像的时间信息、空间信息、图像特征信息， 确定所述第一图像及所述第二图像中分别包含的待确定目标的初步相同概率值；
    所述生成模块，包括：
    第一生成子模块，配置为当所述初步相同概率值大于预设值时，基于所述第一图像及所述第二图像，生成所述预测路径。
  17. 根据权利要求16所述的装置，其中，所述第二确定模块，包括：
    第一确定子模块，配置为将所述第一图像和所述第二图像，以及，所述第一图像和所述第二图像之间的时间信息的差异和空间信息的差异输入孪生卷积神经网络Siamese-CNN，获得所述第一图像和第二图像中的待确定目标的初步相同概率值。
  18. 根据权利要求14-17任一项所述的装置，其中，所述生成模块，包括：
    第二生成子模块，配置为根据所述第一图像的特征信息、所述第一图像的时间信息、所述第一图像的空间信息、所述第二图像的特征信息、所述第二图像的时间信息、和所述第二图像的空间信息，通过概率模型生成所述待确定目标的预测路径。
  19. 根据权利要求18所述的装置，其中，所述第二生成子模块，包括：
    第一确定单元，配置为通过链状马尔可夫随机场模型MRF从获取的图像集中确定出包含有所述待确定目标的信息、且与所述第一图像和所述第二图像均具有时空序列关系的所有图像；
    第一生成单元，配置为根据确定出的所有图像对应的时间信息和空间信息，生成所述待确定目标的预测路径。
  20. 根据权利要求19所述的装置，其中，所述第一生成单元，包括：
    第二生成单元，配置为根据确定出的所有图像对应的时间信息和空间信息，生成以所述第一图像为首节点且以所述第二图像为尾节点的一条预测路径，其中，所述预测路径除对应所述首节点和所述尾节点外，还对应至少一个中间节点。
  21. 根据权利要求20所述的装置，其中，所述第一确定单元，配置为：
    以所述第一图像的空间信息对应的位置为起始位置，以所述第二图像的空间信息对应的位置为终止位置，获取从所述起始位置至所述终止位置之间的所有摄像设备的位置信息；
    根据所有摄像设备的位置信息所指示的位置之间的关系，以所述起始位置对应的摄像设备为起点，以所述终止位置对应的摄像设备为终点，生成至少一个设备路径，其中，每个设备路径除包括所述起点的摄像设备和所述终点的摄像设备外，还包括至少一个其它摄像设备的信息；
    针对每一个设备路径，以所述第一图像的时间信息对应的时间为起始时间，以所述第二图像的时间信息对应的时间为终止时间，从当前路径上的每个其它摄像设备拍摄的图像中，确定出与当前摄像设备相邻的前一摄 像设备拍摄的包含所述待确定目标的信息的图像具有设定时间序列关系、且包含所述待确定目标的信息的图像。
  22. 根据权利要求21所述的装置，其中，所述第二生成单元，配置为：
    针对每一个设备路径，根据确定出的所述图像的时间序列关系生成相连的具有时空序列关系的多个中间节点；根据所述首节点、所述尾节点、和所述中间节点，生成与当前设备路径对应的具有时空序列关系的图像路径；
    从每一个设备路径对应的图像路径中，确定出以所述第一图像为首节点且以所述第二图像为尾节点的最大概率图像路径作为所述待确定目标的预测路径。
  23. 根据权利要求22所述的装置，其中，所述第二生成单元，还配置为：
    针对每一个设备路径对应的图像路径，获取所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率；
    根据所述图像路径中每两个相邻节点的图像之间具有同一个待确定目标的信息的概率计算所述图像路径作为所述待确定目标的预测路径的概率；
    根据每一个图像路径作为所述待确定目标的预测路径的概率确定最大概率图像路径作为所述待确定目标的预测路径。
  24. 根据权利要求14-23任一项所述的装置，其中，所述第一确定模块，包括：
    第二确定子模块，配置为通过神经网络，对所述预测路径进行有效性判断，基于判断结果，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
  25. 根据权利要求24所述的装置，其中，所述第二确定子模块，包括：
    第一获取单元，配置为根据所述预测路径中的相邻的图像的时间信息，获取相邻的图像的时间差异；根据相邻的图像的空间信息，获取相邻的图像的空间差异；根据相邻的图像中的待确定目标的特征信息，获取相邻的图像中的待确定目标的特征差异；
    第二获取单元，配置为将获得的所述预测路径中相邻的图像的时间差异、空间差异和特征差异输入长短时记忆网络LSTM，获得所述预测路径的待确定目标的识别概率；
    第二确定单元，配置为根据所述预测路径的待确定目标的识别概率，确定所述第一图像和第二图像中的待确定目标是否为同一个待确定目标。
  26. 根据权利要求25所述的装置，其中，所述第一获取单元，配置为：
    通过Siamese-CNN分别获取相邻的图像中的待确定目标的特征信息；
    根据分别获取的所述特征信息，获取相邻的图像中的待确定目标的特征差异。
  27. 一种计算机可读存储介质，其上存储有计算机程序指令，其中，所述程序指令被处理器执行时实现权利要求1至13中任意一项权利要求所述的目标识别方法的步骤。
  28. 一种电子设备，包括：处理器、存储器、通信元件和通信总线，所述处理器、所述存储器和所述通信元件通过所述通信总线完成相互间的通信；
    所述存储器用于存放至少一可执行指令，所述可执行指令使所述处理器执行如权利要求1至13中任意一项权利要求所述的目标识别方法的步骤。
PCT/CN2018/097374 2017-07-28 2018-07-27 目标识别方法、装置、存储介质和电子设备 WO2019020103A1 (zh)

Priority Applications (6)

Application Number Priority Date Filing Date Title
SG11201911625YA SG11201911625YA (en) 2017-07-28 2018-07-27 Target recognition method and apparatus, storage medium, and electronic device
JP2019557616A JP6893564B2 (ja) 2017-07-28 2018-07-27 ターゲット識別方法、装置、記憶媒体および電子機器
KR1020197031657A KR102339323B1 (ko) 2017-07-28 2018-07-27 타겟 인식 방법, 장치, 저장 매체 및 전자 기기
US16/565,069 US11200682B2 (en) 2017-07-28 2019-09-09 Target recognition method and apparatus, storage medium, and electronic device
US17/452,776 US20220051417A1 (en) 2017-07-28 2021-10-29 Target recognition method and appartus, storage medium, and electronic device
US17/453,487 US20220058812A1 (en) 2017-07-28 2021-11-03 Target recognition method and appartus, storage medium, and electronic device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710633604.3 2017-07-28
CN201710633604.3A CN108229292A (zh) 2017-07-28 2017-07-28 目标识别方法、装置、存储介质和电子设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/565,069 Continuation US11200682B2 (en) 2017-07-28 2019-09-09 Target recognition method and apparatus, storage medium, and electronic device

Publications (1)

Publication Number Publication Date
WO2019020103A1 true WO2019020103A1 (zh) 2019-01-31

Family

ID=62654256

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/097374 WO2019020103A1 (zh) 2017-07-28 2018-07-27 目标识别方法、装置、存储介质和电子设备

Country Status (6)

Country Link
US (3) US11200682B2 (zh)
JP (1) JP6893564B2 (zh)
KR (1) KR102339323B1 (zh)
CN (1) CN108229292A (zh)
SG (1) SG11201911625YA (zh)
WO (1) WO2019020103A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021017891A1 (zh) * 2019-07-31 2021-02-04 腾讯科技（深圳）有限公司 对象跟踪方法和装置、存储介质及电子设备
CN112766301A (zh) * 2020-12-11 2021-05-07 南京富岛信息工程有限公司 一种采油机示功图相似性判断方法
JP2022539250A (ja) * 2019-07-03 2022-09-07 ウェイモ エルエルシー アンカー軌道を使用したエージェント軌道予測

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229292A (zh) 2017-07-28 2018-06-29 北京市商汤科技开发有限公司 目标识别方法、装置、存储介质和电子设备
CN108921811B (zh) * 2018-04-03 2020-06-30 阿里巴巴集团控股有限公司 检测物品损伤的方法和装置、物品损伤检测器
CN109508787A (zh) * 2018-10-16 2019-03-22 深圳大学 用于超声位移估计的神经网络模型训练方法及系统
CN111160067A (zh) * 2018-11-07 2020-05-15 北京奇虎科技有限公司 危险识别方法、装置、电子设备及计算机可读存储介质
CN109740479A (zh) * 2018-12-25 2019-05-10 苏州科达科技股份有限公司 一种车辆重识别方法、装置、设备及可读存储介质
CN109726684B (zh) * 2018-12-29 2021-02-19 百度在线网络技术（北京）有限公司 一种地标元素获取方法和地标元素获取系统
US10373323B1 (en) * 2019-01-29 2019-08-06 StradVision, Inc. Method and device for merging object detection information detected by each of object detectors corresponding to each camera nearby for the purpose of collaborative driving by using V2X-enabled applications, sensor fusion via multiple vehicles
CN110490906A (zh) * 2019-08-20 2019-11-22 南京邮电大学 一种基于孪生卷积网络和长短期记忆网络的实时视觉目标跟踪方法
CN110991413B (zh) * 2019-12-20 2020-12-15 西南交通大学 一种基于ReID的跑步检测方法
CN113627260A (zh) * 2021-07-12 2021-11-09 科大讯飞股份有限公司 识别手写汉字的笔顺的方法、系统和计算设备
CN113673412B (zh) * 2021-08-17 2023-09-26 驭势（上海）汽车科技有限公司 关键目标物的识别方法、装置、计算机设备及存储介质
CN113688776B (zh) * 2021-09-06 2023-10-20 北京航空航天大学 一种用于跨视场目标重识别的时空约束模型构建方法
CN114338974A (zh) * 2021-12-02 2022-04-12 深圳市领航卫士安全技术有限公司 多通道的活动路径确定方法、装置、设备及存储介质
CN114783003B (zh) * 2022-06-23 2022-09-20 之江实验室 一种基于局部特征注意力的行人重识别方法和装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105095362A (zh) * 2015-06-25 2015-11-25 深圳码隆科技有限公司 一种基于目标对象的图像显示方法和装置
CN106326837A (zh) * 2016-08-09 2017-01-11 北京旷视科技有限公司 对象追踪方法和装置
CN106650660A (zh) * 2016-12-19 2017-05-10 深圳市华尊科技股份有限公司 一种车型识别方法及终端
CN108229292A (zh) * 2017-07-28 2018-06-29 北京市商汤科技开发有限公司 目标识别方法、装置、存储介质和电子设备

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001208844A (ja) * 2000-01-25 2001-08-03 Sumitomo Electric Ind Ltd 光学式車両感知装置および光学式車両感知方法
US7623674B2 (en) * 2003-11-05 2009-11-24 Cognex Technology And Investment Corporation Method and system for enhanced portal security through stereoscopy
JP4447309B2 (ja) * 2003-12-25 2010-04-07 財団法人生産技術研究奨励会 交差点交通量計測装置および交差点交通量計測方法
US9002060B2 (en) * 2012-06-28 2015-04-07 International Business Machines Corporation Object retrieval in video data using complementary detectors
WO2014024264A1 (ja) * 2012-08-08 2014-02-13 株式会社 日立製作所 交通量予測装置および方法
US9436895B1 (en) * 2015-04-03 2016-09-06 Mitsubishi Electric Research Laboratories, Inc. Method for determining similarity of objects represented in images
JP6439571B2 (ja) * 2015-04-28 2018-12-19 オムロン株式会社 交通情報収集装置、交通情報収集方法、および交通情報収集プログラム
JP6433877B2 (ja) * 2015-10-27 2018-12-05 日本電信電話株式会社 目的地予測装置、目的地予測方法、及び目的地予測プログラム
JP6521835B2 (ja) * 2015-10-27 2019-05-29 日本電信電話株式会社 移動経路予測装置、移動経路予測方法、及び移動経路予測プログラム
EP3403216B1 (en) * 2016-01-11 2023-11-01 Mobileye Vision Technologies Ltd. Systems and methods for augmenting upright object detection
CN106778517A (zh) * 2016-11-25 2017-05-31 河南高速公路驻信段改扩建工程有限公司 一种监控视频序列图像车辆再识别的方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105095362A (zh) * 2015-06-25 2015-11-25 深圳码隆科技有限公司 一种基于目标对象的图像显示方法和装置
CN106326837A (zh) * 2016-08-09 2017-01-11 北京旷视科技有限公司 对象追踪方法和装置
CN106650660A (zh) * 2016-12-19 2017-05-10 深圳市华尊科技股份有限公司 一种车型识别方法及终端
CN108229292A (zh) * 2017-07-28 2018-06-29 北京市商汤科技开发有限公司 目标识别方法、装置、存储介质和电子设备

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2022539250A (ja) * 2019-07-03 2022-09-07 ウェイモ エルエルシー アンカー軌道を使用したエージェント軌道予測
JP7459224B2 (ja) 2019-07-03 2024-04-01 ウェイモ エルエルシー アンカー軌道を使用したエージェント軌道予測
WO2021017891A1 (zh) * 2019-07-31 2021-02-04 腾讯科技（深圳）有限公司 对象跟踪方法和装置、存储介质及电子设备
CN112766301A (zh) * 2020-12-11 2021-05-07 南京富岛信息工程有限公司 一种采油机示功图相似性判断方法
CN112766301B (zh) * 2020-12-11 2024-04-12 南京富岛信息工程有限公司 一种采油机示功图相似性判断方法

Also Published As

Publication number Publication date
US20220051417A1 (en) 2022-02-17
SG11201911625YA (en) 2020-01-30
KR102339323B1 (ko) 2021-12-14
US11200682B2 (en) 2021-12-14
JP2020519989A (ja) 2020-07-02
CN108229292A (zh) 2018-06-29
JP6893564B2 (ja) 2021-06-23
US20200005090A1 (en) 2020-01-02
US20220058812A1 (en) 2022-02-24
KR20190128724A (ko) 2019-11-18

Similar Documents

Publication Publication Date Title
WO2019020103A1 (zh) 目标识别方法、装置、存储介质和电子设备
US11643076B2 (en) Forward collision control method and apparatus, electronic device, program, and medium
US11182592B2 (en) Target object recognition method and apparatus, storage medium, and electronic device
US10599958B2 (en) Method and system for classifying an object-of-interest using an artificial neural network
CN107851318B (zh) 用于对象跟踪的系统和方法
CN109035304B (zh) 目标跟踪方法、介质、计算设备和装置
US11600008B2 (en) Human-tracking methods, systems, and storage media
US9760800B2 (en) Method and system to detect objects using block based histogram of oriented gradients
CN110765860A (zh) 摔倒判定方法、装置、计算机设备及存储介质
WO2020047854A1 (en) Detecting objects in video frames using similarity detectors
CN110660102B (zh) 基于人工智能的说话人识别方法及装置、系统
US20190171910A1 (en) Best Image Crop Selection
WO2016145591A1 (en) Moving object detection based on motion blur
KR20210012012A (ko) 물체 추적 방법들 및 장치들, 전자 디바이스들 및 저장 매체
WO2015186347A1 (ja) 検出システム、検出方法及びプログラム記憶媒体
US10599946B2 (en) System and method for detecting change using ontology based saliency
CN111079621A (zh) 检测对象的方法、装置、电子设备和存储介质
CN114663871A (zh) 图像识别方法、训练方法、装置、系统及存储介质
JPWO2018179119A1 (ja) 映像解析装置、映像解析方法およびプログラム
CN111310595A (zh) 用于生成信息的方法和装置
CN112819859B (zh) 一种应用于智慧安防的多目标跟踪方法及装置
US9842406B2 (en) System and method for determining colors of foreground, and computer readable recording medium therefor
CN112347853A (zh) 一种基于视频的车牌数据脱敏方法、存储介质及服务器
CN111860070A (zh) 识别发生改变的对象的方法和装置
CN114596580B (zh) 一种多人体目标识别方法、系统、设备及介质

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: 18838049

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019557616

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20197031657

Country of ref document: KR

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18838049

Country of ref document: EP

Kind code of ref document: A1