CN112418262A - Vehicle re-identification method, client and system - Google Patents

Vehicle re-identification method, client and system Download PDF

Info

Publication number
CN112418262A
CN112418262A CN202011011077.0A CN202011011077A CN112418262A CN 112418262 A CN112418262 A CN 112418262A CN 202011011077 A CN202011011077 A CN 202011011077A CN 112418262 A CN112418262 A CN 112418262A
Authority
CN
China
Prior art keywords
vehicle
feature
image
distance
images
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202011011077.0A
Other languages
Chinese (zh)
Inventor
王茜
刘民
张鸿洲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
SHANGHAI CRIMINAL SCIENCE TECHNOLOGY RESEARCH INSTITUTE
Original Assignee
SHANGHAI CRIMINAL SCIENCE TECHNOLOGY RESEARCH INSTITUTE
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 SHANGHAI CRIMINAL SCIENCE TECHNOLOGY RESEARCH INSTITUTE filed Critical SHANGHAI CRIMINAL SCIENCE TECHNOLOGY RESEARCH INSTITUTE
Priority to CN202011011077.0A priority Critical patent/CN112418262A/en
Publication of CN112418262A publication Critical patent/CN112418262A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles

Abstract

The invention discloses a vehicle re-identification method, a client and a system, and relates to the technical field of computer vision and pattern recognition. The method comprises the following training process: inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model; and (3) re-identification process: acquiring a query vehicle image input by a user, respectively acquiring the feature vectors of the query vehicle image and each image in a comparison library based on a basic network model, calculating the distance between the feature vectors by constructing a depth measurement learning network, implementing sequencing according to the distance and outputting an identification result; the feature vector is a global feature generated by adding a plurality of local block features fused into the view angle feature and the relative position information and then performing feature fusion. The invention integrates the multi-attribute characteristics, increases the position information corresponding to the characteristics, can adapt to the change of the visual angle, and optimizes the identification rate and the cost performance of the calculated amount.

Description

Vehicle re-identification method, client and system
Technical Field
The invention relates to the technical field of computer vision and pattern recognition, in particular to a vehicle re-recognition method, a client and a system.
Background
With the widespread application of monitoring cameras in the field of public safety, vehicles are used as important objects in urban monitoring, and the vehicles attract wide attention in a large number of tasks related to the vehicles, such as detection, tracking, classification, verification and the like. Vehicle re-identification is to find out vehicles that contain images captured by other cameras of the same vehicle or under different lighting, viewing angles, etc. under the same camera as the query image. Through vehicle re-identification, a target vehicle can automatically find, locate and track across a plurality of cameras, and the method plays an important role in the automatic analysis of increasingly-increased urban monitoring videos. The application scenarios of vehicle re-identification are very wide, such as vehicle tracking, vehicle positioning, criminal detection and the like, and are also very important for the application of intelligent transportation and smart cities. The vehicle image shot by the monitoring system without the overlapped vision field is a main processing object for vehicle re-identification. Because the shot vehicle image contains factors such as visual angle (position) change, resolution, illumination change, blurring, camera setting, complex background and shielding, the vehicle re-identification problem has special difficulties, and solutions to the difficulties are still researched by many scholars.
Currently, methods for vehicle re-identification under non-overlapping field-of-view monitoring systems can be broadly divided into two categories: one is a vehicle re-identification method based on feature learning, and the other is a vehicle re-identification algorithm based on metric learning. For the vehicle re-identification method based on the feature learning, the vehicle image is represented as a feature vector by some descriptors with robustness, stability and distinguishing capability. In the vehicle re-identification problem, the similarity between vehicle images is directly or indirectly measured by extracting descriptive features which are distinctive to different vehicles and robust to view and illumination. Commonly used target re-identification features typically include color histograms (such as HSV, RGB, YCbCr color histogram), texture histograms, Local Binary Patterns (LBP), Scale Invariant Feature Transforms (SIFT), and the like. However, since the feature matching has no good robustness to the illumination and the change of the view angle (perspective), it is difficult to satisfy the requirements of robustness, stability and distinction at the same time. For a metric learning based vehicle re-identification method, it translates the vehicle re-identification problem into a distance metric problem between the computed target image and the candidate image. Distances commonly used for metrics typically include euclidean distance, mahalanobis distance, cosine distance, hamming distance, and the like. In recent years, deep metric learning is also a common method in vehicle re-identification problems, for example, training a multi-layer convolutional neural network to simultaneously extract features of vehicle images and corresponding similarity metric function learning. Vehicle re-identification based on depth metric learning does not need to manually extract the characteristics of a vehicle, but trains a large number of vehicle images by applying a neural network model. However, the training complexity is large, for example, an imbalance problem caused by training an imbalance data set is caused, and for image databases with different magnitudes, due to the imbalance of the image data set, there are often many more unmatched pairs than matched pairs.
In summary, the conventional vehicle re-identification method is difficult to achieve in the aspects of adaptive view angle identification, data imbalance, identification rate and computational cost ratio, and how to provide a vehicle re-identification scheme which can adapt to view angle changes, reduce data imbalance influence and has superior identification rate and computational cost ratio based on the prior art is a technical problem which needs to be solved urgently.
Disclosure of Invention
The invention aims to: the defects of the prior art are overcome, and the vehicle re-identification method, the client and the system are provided. The vehicle re-identification method provided by the invention integrates the multi-attribute characteristics, increases the corresponding position information of the characteristics, can adapt to the change of the visual angle, and optimizes the identification rate and the cost performance of the calculated amount.
In order to achieve the above object, the present invention provides the following technical solutions:
a vehicle re-identification method comprises a training process and a re-identification process, and comprises the following steps:
training process: inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model;
and (3) re-identification process: acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing a depth measurement learning network, implementing sequencing according to the distance, and outputting an identification result;
the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion.
Further, the training process may be preceded by the step of constructing a set of vehicle images, as follows:
erecting a plurality of camera terminal groups in a real lane, wherein each camera terminal group comprises a plurality of camera terminal devices to form a group of linked camera groups of 0-180 degrees around a vehicle; each group of camera terminal groups comprises 7 camera terminal devices, and the visual angle is divided into 7 visual angles which are respectively 0 degree, 30 degrees, 60 degrees, 90 degrees, 120 degrees, 150 degrees and 180 degrees;
the method comprises the steps that a plurality of groups of vehicle video images and/or vehicle static images at a plurality of visual angles are collected through the camera terminal equipment;
and screening the images, and selecting a clear picture corresponding to each view angle of each vehicle to be selected into a vehicle image set.
Further, the training process comprises a feature extraction step, a metric learning step and a classification step;
the characteristic extraction step comprises the steps of obtaining each image and extracting a plurality of local block characteristics of the vehicle in the image;
the metric learning step comprises the steps of converting the local block features into local block features fused with view angle features through a plurality of set basic feature views, carrying out fusion calculation on the local block features fused with the view angle features to obtain global features, and carrying out distance metric learning on the global features.
Further, the local blocking bitsCharacterizing a color histogram feature f comprising a host vehicle with a relative position dimensioncrFull pixel color histogram feature fchAnd texture feature fvThe feature extraction step is as follows:
vehicle image set with aspect ratio W × D for image set
Figure BDA0002697569290000039
Two different vehicles are taken from the training set and are recorded as M and N, M,
Figure BDA0002697569290000031
the test vehicle is noted as T,
Figure BDA0002697569290000032
extracting a color component f of the host vehicle with a relative position dimension for each vehicle imagecrThe following are:
dividing the vehicle image into 12 x 12 non-overlapping blocks Pw，dAnd acquiring a block color histogram and recording the block color histogram as:
Figure BDA0002697569290000033
wherein K represents the number of dereferencing characteristics, and K represents the number of dereferencing characteristics; h (k) is the number of pixels in the image having a characteristic value of k; w and d are respectively the central horizontal and longitudinal axis coordinates of the block;
for dimension reduction and improvement of separability, the component values of the first 10% of h (k) in h (k) represent the block color characteristics and are recorded as
Figure BDA0002697569290000034
Meanwhile, the center position of the score block is (c)w，cd) The center position of the block distance graph is
Figure BDA0002697569290000035
Obtaining the block color characteristics containing the block position information as follows:
Figure BDA0002697569290000036
wherein pw and pd respectively represent the row and column numbers of the block, and fcrThe calculation formula of (a) is as follows:
Figure BDA0002697569290000037
where, s is 1, 2, …, pw × pd.
Based on formula (1), full-pixel color histogram information f is acquiredchThe following are:
Figure BDA0002697569290000038
obtaining texture features by extracting FV coding dense features as follows:
converting the vehicle image into a gray image, dividing the vehicle image into 24 × 24 non-overlapping blocks P for the vehicle images of multiple view anglesw′，d′Extracting low-level SIFT descriptors point by point from each 24 × 24 image block; w ', d ' are respectively the central horizontal and vertical axis coordinates of the block, and the central position of the block is marked as (c 'w，c′d)；
SIFT descriptor is obtained by using PCA algorithm to reduce dimension to l-128 dimension
Figure BDA0002697569290000041
And adding the central point position to obtain the block color characteristics containing the block position information, as follows:
Figure BDA0002697569290000042
v to be extractedpatchPerforming feature fitting to obtain a Gaussian mixture model GMM, wherein the model comprises H Gaussian models, and enabling omega to be omegah，γh，ShRespectively representing the weight, mean and covariance matrix of the H model in the GMM, H is 0, 1, …, H-1; calculating each Gaussian model according to the following formulaFirst order differential v of the form(1)And a second order differential v(2)：
Figure BDA0002697569290000043
Figure BDA0002697569290000044
Wherein N represents the number of features;
Figure BDA00026975692900000411
representing the weight of the nth feature for the h gaussian model, N being 0, 1, …, N-1; x is the number ofnA feature value representing an nth feature;
will be provided with
Figure BDA0002697569290000045
And
Figure BDA0002697569290000046
are connected in series to obtain:
Figure BDA0002697569290000047
and adding position information to obtain (l +2) × H × 2 dimensional integral FV characteristics, wherein l is 128, and the FV characteristics are as follows:
Figure BDA0002697569290000048
sort (-) function sorting is carried out on the energy, Z key positions capable of representing the textural features of the energy are selected to form fvThe expression is as follows:
Figure BDA0002697569290000049
wherein ss is 1, 2, …, Z, Z is an integer greater than 1.
Further, setting 0, 90 and 180 degrees as three basic feature perspectives, and calculating to obtain the global features according to the following steps:
marking three basic characteristic visual angles i of 0 degree, 90 degrees and 180 degrees as V respectivelyfro、Vsid、VrevConverting all images in the training set into a picture characteristic function v blended with the visual angle characteristicsnullThe following are:
Figure BDA00026975692900000410
wherein alpha isi，θA view transformation characteristic function representing a view i;
the method for obtaining the characteristics of the blocks FV according to the formula (9) and the formula (11), the formula (9) can be expressed as the characteristics blended into the characteristics of the view angle:
Figure BDA0002697569290000051
is calculated to obtain
Figure BDA0002697569290000052
Simultaneously obtain f'ch，
Figure BDA0002697569290000053
Formula (3) is similarly rewritten and expressed as a feature that has been incorporated into the viewing angle feature as follows:
Figure BDA0002697569290000054
calculating to obtain a global feature FfullThe following are:
Ffull＝β1，θf′cr+β2，θf′ch+β3，θf′v (14)
wherein, betaj，θRepresenting the weight value of each characteristic component according to the change of the visual angle, j is more than or equal to 1 and less than or equal toAnd 3, classification feature classes.
Further, the distance metric learning adopts mahalanobis distance, and for any two vehicle images, the step of calculating the distance function between the feature vectors of the two vehicle images through the mahalanobis distance is as follows:
for any two vehicle images si and sj, a distance function d (y) based on a characteristic vector between the images si and sj is calculated through a horse-type distancesi，ysj) The following are:
Figure BDA0002697569290000055
wherein, A is a diagonal matrix which represents the weight to distinguish the importance degree of the characteristic vector; decomposing A into WTW is solved instead of the weight beta in equation (14)j，θThe function of (1); z is a radical ofsiIs the feature vector of si, zsiA feature vector of sj;
through rijE (1, -1) to express whether the two images are similar, and the similarity belongs to the same class, rij1, otherwise, r is not of the same classijThe constraint function is given by-1:
Figure BDA0002697569290000056
wherein the content of the first and second substances,
Figure BDA0002697569290000057
the critical value parameter represents the distance between classes and the class distance, delta (si, sj) represents a positive and negative sample equalization function, and the calculation formula of delta (si, sj) is as follows:
Figure BDA0002697569290000061
wherein z issnThe minimum value of the distance of all negative samples;
and (3) realizing the pairwise metric learning by calculating W, and obtaining final classification information by a KNN method.
Further, in the training process, for a large image database, model training is carried out by adopting a triple depth measurement learning algorithm, wherein the triple comprises a group of test images T and two groups of training images M and N; when the test acquisition test image T is close to M and far away from N, M, N is corresponding to the global feature { F } of Tfull(M)，Ffull(N)，Ffull(T) is adjusted to { y }M，yN，yTAnd (5) rewriting the distance constraint conditions of the feature vectors of the three components according to the formula (15) as follows:
Figure BDA0002697569290000062
wherein the content of the first and second substances,
Figure BDA0002697569290000065
is a critical value parameter of the distance between classes and the class distance; zT、ZN、ZMFeature vectors, z, for images T, M and N, respectivelysiIs the feature vector of sj.
Further, in the training process, aiming at the pair method, the asymmetric problem is solved by adopting a weight strategy of an asymmetric sample, wherein the weight strategy of the asymmetric sample is as follows: once per iteration, num + +, by weight parameter
Figure BDA0002697569290000063
Adjusting the weight value;
aiming at the triple depth metric learning algorithm, in order to reduce the operand, only triplets which violate the constraint condition and reach the non-zero loss constraint condition are calculated; and the number of the first and second groups,
for those in the formula (18)
Figure BDA0002697569290000064
The values are sorted from large to small, and sample matching is performed from the maximum value to reduce the interference of noise images.
The invention also provides a vehicle re-identification client, which comprises the following structure:
the model training module is used for inputting a vehicle image set containing different vehicle images into the convolutional neural network model as input data to train until the model converges to obtain a basic network model;
and a re-identification module: the system comprises a base network model, a comparison library, a query vehicle image and a depth measurement learning network, wherein the base network model is used for acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and the comparison library based on the base network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing the depth measurement learning network, implementing sequencing according to the distance and outputting an identification result; the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion. The invention also provides a vehicle re-identification system, which comprises a user terminal and a server,
the user terminal is provided with a human-computer interaction interface, and a vehicle image set for training and query information input by a user are acquired through the human-computer interaction interface;
the server side includes a processor and a memory for storing processor-executable instructions and parameters, the processor configured to:
inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model; acquiring a query vehicle image input by a user, respectively acquiring the feature vector of each image in the query vehicle image and the comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing a depth measurement learning network, implementing sequencing according to the distance, and outputting an identification result; the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion. Due to the adoption of the technical scheme, compared with the prior art, the visual angle display device has the following advantages and positive effects as examples:
aiming at the characteristics of a vehicle image set, a vehicle re-identification algorithm which is based on fusion characteristics and metric learning and can adapt to visual angle change is realized, links such as fusion of multi-attribute characteristics, increase of position information corresponding to the characteristics, coping with data unbalance, identification rate and calculation capacity cost ratio selection schemes are improved, and the identification rate is improved by testing on image data sets with different magnitudes and a public image data set. Further, in a metric learning link of model training, an improved pair metric learning method with high recognition rate and high computation cost performance and a triple depth metric learning (triple deep metric learning) algorithm with high recognition rate combined with a CNN network are respectively selected, and the asymmetry problem is solved by setting a weight strategy aiming at the asymmetry problem of image sets (samples) with different magnitudes.
On one hand, the method expands 5 visual angles of the previous multi-visual-angle identification database building into 7 visual angles, and carries out basic image database collection and modeling according to the secondary building real-time linkage hardware acquisition scheme. Meanwhile, relative position information acquired during feature extraction is added as dimension features of metric learning, any view angle vehicle is defined as a feature transformation function based on three basic components of 0 degrees and 90 degrees and 180 degrees, a block mapping scheme is implemented for each primary feature component, a training set is trained, a composite feature function capable of self-recognition and self-adaptive view angle change is acquired, and the recognition effect is improved.
On the other hand, different solutions for different algorithms are adopted according to the data volume property of the image set for training the unbalanced image library, the medium-sized database can adopt a pairwise method, and the large-sized database can adopt a triplets method. And meanwhile, the composite distance values are sequenced and then sample matching is carried out from the maximum value so as to reduce the interference of noise images.
Drawings
Fig. 1 is an information processing diagram of a vehicle re-identification method according to an embodiment of the present invention.
Fig. 2 is a schematic flow chart of vehicle re-identification according to an embodiment of the present invention.
Fig. 3 is a block diagram of a client according to an embodiment of the present invention.
Fig. 4 is a block diagram of a system according to an embodiment of the present invention.
Description of reference numerals:
a client 200, a model training module 210, and a re-identification module 220;
system 300, user terminal 310, server 320.
Detailed Description
The method, client and system for vehicle re-identification disclosed in the present invention are further described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that technical features or combinations of technical features described in the following embodiments should not be considered as being isolated, and they may be combined with each other to achieve better technical effects. In the drawings of the embodiments described below, the same reference numerals appearing in the respective drawings denote the same features or components, and may be applied to different embodiments. Thus, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.
It should be noted that the structures, proportions, sizes, and other dimensions shown in the drawings and described in the specification are only for the purpose of understanding and reading the present disclosure, and are not intended to limit the scope of the invention, which is defined by the claims, and any modifications of the structures, changes in the proportions and adjustments of the sizes and other dimensions, should be construed as falling within the scope of the invention unless the function and objectives of the invention are affected. The scope of the preferred embodiments of the present invention includes additional implementations in which functions may be executed out of order from that described or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the embodiments of the present invention.
Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate. In all examples shown and discussed herein, any particular value should be construed as merely illustrative, and not limiting. Thus, other examples of the exemplary embodiments may have different values.
Examples
Referring to fig. 1, a vehicle re-identification method provided by the present invention includes the following steps:
step S100 training process: and (3) taking a vehicle image set containing different vehicle images as input data, inputting the input data into the convolutional neural network model, and training until the model converges to obtain a basic network model.
Step S200 re-identification process: acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing a depth measurement learning network, implementing sequencing according to the distance, and outputting an identification result; the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion.
Preferably, the training process further comprises a step of constructing a vehicle image set, and the specific steps are as follows:
firstly, erecting a plurality of camera terminal groups in a real lane, wherein each camera terminal group comprises a plurality of camera terminal devices to form a 0-180-degree linkage camera group around a vehicle. Preferably, each group of camera terminal devices includes 7 camera terminal devices, and the viewing angles are divided into 7 viewing angles, which are 0 °, 30 °, 60 °, 90 °, 120 °, 150 °, and 180 °, respectively.
And then, a plurality of groups of vehicle video images and/or vehicle still images with a plurality of visual angles are collected through the camera terminal equipment.
And then, screening the images, and selecting a clear picture corresponding to each view angle of each vehicle to be selected into a vehicle image set.
In this embodiment, referring to fig. 1, the training process may include three steps of feature extraction (reconstruction), Metric learning (Metric learning), and Classification (Classification), and correspondingly, the obtained basic network model includes a feature extraction module, a distance Metric learning module, and a Classification module. The basic network model may be based on a mainstream neural network structure, including but not limited to google lenet, VGGNet, ResNet, etc., and in this embodiment, a google net V4 network structure is preferred.
The feature extraction step includes acquiring each image, and extracting a plurality of local block features of the vehicle in the image.
By way of example and not limitation, the local blocking features used for identification may be a color histogram (such as HSV, RGB, YCbCr color histogram), a texture histogram, a Local Binary Pattern (LBP), and/or a scale-invariant feature transform (SIFT).
The metric learning step comprises the steps of converting the local block features into local block features fused with view angle features through a plurality of set basic feature views, carrying out fusion calculation on the local block features fused with the view angle features to obtain global features, and carrying out distance metric learning on the global features.
By way of example and not limitation, the distance metric learning may employ Euclidean distance, Mahalanobis distance, cosine distance, and Hamming distance.
Including the color histogram feature f of the host vehicle with the relative position dimension as a local blocking featurecrFull pixel color histogram feature fchAnd texture feature fvA mahalanobis distance is taken as an example of a distance measurement method, and a typical implementation of the present embodiment is described in detail.
A vehicle image set is constructed by sample acquisition, the sample acquisition scheme being as follows:
constructing 42 camera terminal devices to collect vehicle videos and images together, dividing 42 terminals into 6 groups to be erected, constructing one camera terminal at every 30 degrees from the left side of a real lane, totally 7 camera terminals at 0-180 degrees of a group of ring cars, and numbering the camera terminals as Sa，b(a is more than or equal to 1 and less than or equal to 6, and b is more than or equal to 1 and less than or equal to 7), wherein a is a group, and b is the number of different viewing angles in the same group. Collection by Sa，bMultiple vehicle video feed of shootingAnd (4) performing multi-frame screening, and confirming that 7 visual angles of each vehicle are respectively cut out and a clear picture is put in storage. By the hardware erection method, the classification error caused by visual angle labeling during library building can be reduced.
And acquiring the training sample with the marked visual angle classification. Preferably, 6 groups of camera terminals are all arranged on a certain straight road section, so that the condition that a part of cameras do not capture the vehicle can be eliminated. The number of pictures taken by the same vehicle through the erection section at one time is 7 to 42. Considering that long-term data accumulation can cause the imbalance of the vehicle image database samples due to the fact that the number of the same vehicle image can reach up to 200, different algorithms are used according to different conditions when the imbalance database is trained, and estimation error influence caused by the imbalance is effectively reduced, which is described in detail later.
Then, feature extraction, metric learning, and classification are performed as follows.
1) Feature extraction
Let the image set be a vehicle image set with aspect ratio W × D
Figure BDA0002697569290000107
Two different vehicles are taken in the training set and are recorded as M and N
Figure BDA0002697569290000101
Test vehicle noted T
Figure BDA0002697569290000102
As shown in fig. 2. For each vehicle image, first, a color component f of the host vehicle with a relative position dimension is extractedcrThe method comprises the following steps:
dividing the vehicle image into 12 x 12 non-overlapping blocks Pw，dAnd acquiring a block color histogram and recording the block color histogram as:
Figure BDA0002697569290000103
wherein K represents the number of dereferencing characteristics, and K represents the number of dereferencing characteristics; h (k) is the number of pixels in the image having a characteristic value of k; and w and d are respectively the central horizontal and vertical axis coordinates of the block.
In order to reduce dimension (reduce feature dimension) and improve separability, in this embodiment, it is preferable to take the first 10% of h (k) component values in h (k) to represent the block color feature, and record the block color feature as
Figure BDA0002697569290000104
Meanwhile, the center position of the score block is (c)w，cd) The center position of the block distance graph is
Figure BDA0002697569290000105
Obtaining the block color characteristics containing the block position information as follows:
Figure BDA0002697569290000106
where pw and pd represent the number of rows and columns of the block, respectively.
fcrThe calculation formula of (a) is as follows:
Figure BDA0002697569290000111
where, s is 1, 2, …, pw × pd.
Based on the formula (1), the full-pixel color histogram information f can be acquiredchThe following are:
Figure BDA0002697569290000112
preferably, the texture features are obtained by extracting FV (i.e., Fisher vector) encoded dense features as follows:
first, a vehicle image is converted into a grayscale image, and the vehicle image is divided into 24 × 24 non-overlapping blocks P for the vehicle images of a plurality of view anglesw′，d′Low-level SIFT descriptors are extracted point-by-point for each 24 × 24 patch for 7 views. In the above value-taking method, more emphasis is placed on feature selection based on parts or chunks (patch), and the method is suitable for use in parts or chunks (patch)And (4) identifying targets with larger scales on vehicles and the like.
w ', d ' are respectively the central horizontal and vertical axis coordinates of the block, and the central position of the block is marked as (c 'w，c′d). In this embodiment, w', d, c may be set to w ═ d ═ cw′＝cw，cd′＝cd. And then using PCA algorithm to reduce dimension to l-128 dimension to obtain SIFT descriptor
Figure BDA0002697569290000113
And adding the central point position to obtain the block color characteristics containing the block position information, as follows:
Figure BDA0002697569290000114
v to be extractedpatchPerforming feature fitting to obtain a Gaussian mixture model GMM, wherein the model comprises H Gaussian models, and enabling omega to be omegah，γh，ShThe weights, means and covariance matrices of the H-th model in GMM are respectively represented, H is 0, 1, …, H-1.
The first order differential v of each Gaussian model is calculated as follows(1)And a second order differential v(2)：
Figure BDA0002697569290000115
Figure BDA0002697569290000116
Wherein N represents the number of features;
Figure BDA0002697569290000117
well show the weight of the nth feature for the h gaussian model, N ═ 0, 1, …, N-1; x is the number ofnA feature value representing an nth feature;
will be provided with
Figure BDA0002697569290000118
And
Figure BDA0002697569290000119
are connected in series to obtain:
Figure BDA00026975692900001110
and adding position information to obtain (l +2) × H × 2 dimensional integral FV characteristics, wherein l is 128, and the FV characteristics are as follows:
Figure BDA0002697569290000121
sort (-) function sorting is carried out on the energy, Z key positions capable of representing the textural features of the energy are selected to form fvThe expression is as follows:
Figure BDA0002697569290000122
2) metric learning incorporating perspective transformation
Wherein ss is 1, 2, …, Z, Z is an integer greater than 1. In this example, Z is taken as 100.
Three basic characteristic visual angles are set as 0, 90 and 180 degrees.
Marking three basic characteristic visual angles i of 0 degree, 90 degrees and 180 degrees as V respectivelyfro、Vsid、VrevTraining set
Figure BDA0002697569290000128
All images in the picture can be converted into a picture characteristic function v with visual angle characteristics blended innullThe following are:
Figure BDA0002697569290000123
wherein alpha isi，θA view transformation characteristic function representing the view i. In this example, αi，θThe value may be 0.
The method of finding the features of the patches FV according to equation (9) (by default each patch is not divided and only a single base feature is mapped)) and equation (11) can express equation (9) as a feature that has been blended into the view-angle feature:
Figure BDA0002697569290000124
is calculated to obtain
Figure BDA0002697569290000125
Simultaneously obtain f'ch，
Figure BDA0002697569290000126
Formula (3) is similarly rewritten and expressed as a feature that has been incorporated into the viewing angle feature as follows:
Figure BDA0002697569290000127
finally, global feature F is obtained through calculationfullThe following are:
Ffull＝β1，θf′cr+β2，θf′ch+β3，θf′v (14)
wherein, betaj，θAnd j is more than or equal to 1 and less than or equal to 3, and is used as a classification characteristic category.
The global feature is formed by adding a plurality of local block features blended into the view angle feature and then performing feature fusion after adding relative position information, and may also be referred to as a fusion feature, as shown in fig. 2, after the fusion feature is generated, a depth metric network (or referred to as a depth metric learning network) may be constructed.
In this embodiment, for any two vehicle images in the image set, for any two vehicle images si and sj, a distance function d (y) based on a feature vector between the images si and sj is calculated by a horse-like distancesi，ysj) As follows：
Figure BDA0002697569290000131
Where A is a diagonal matrix representing weights to distinguish the importance of each attribute, and is decomposed into WTW is solved to replace the weight beta in (14)j，θThe function of (1); z is a radical ofsiIs a feature vector of the vehicle image si, zsiIs the feature vector of the vehicle image sj.
In this embodiment, the constraint function can be obtained by expressing whether the two categories belong to the same category (the same category is 1, and vice versa — 1).
Specifically, by rijE (1, -1) to express whether the two images are similar, and the similarity belongs to the same class, rij1, otherwise, r is not of the same classijThe constraint function is given by-1:
Figure BDA0002697569290000132
wherein the content of the first and second substances,
Figure BDA0002697569290000133
the critical value parameter represents the distance between classes and the class distance, delta (si, sj) represents a positive and negative sample equalization function, and the calculation formula of delta (si, sj) is as follows:
Figure BDA0002697569290000134
where zsn is the minimum distance for all negative examples.
Finally, pair metric learning can be achieved by solving W, and final classification information is obtained by a KNN (i.e., K Nearest Neighbor) method.
In this embodiment, to adapt to different data sets, it is preferable to adopt a pairwise method for the medium-size image database, and it is preferable to adopt a triplets (triple) method for the large-size image database.
Specifically, a triple deep metric learning (triplet deep learning) algorithm is adopted for model training of a large image database.
The triplet includes a set of test images T and two sets of training images M and N. When the input test image T is close to M and far away from N, M, N will correspond to the global feature { F } of Tfull(M)，Ffull(N)，Ffull(T) is adjusted to { y }M，yN，yT}. Then, the distance constraint of the feature vectors of the three is rewritten according to the formula (15) as follows:
Figure BDA0002697569290000141
wherein the content of the first and second substances,
Figure BDA0002697569290000142
is a critical value parameter of the distance between classes and the class distance; zT、ZN、ZMFeature vectors, z, for images T, M and N, respectivelysiIs the feature vector of sj.
And finally, selecting GoogleNet V4 as a CNN network to complete the triplet deep metric learning algorithm.
According to the data volume property of the image data set, the method adopts the formula (16) and the formula (18) respectively, and effectively solves the problem caused by the imbalance of the database.
In this embodiment, for the pair wise method, in addition to setting a δ (si, sj) positive and negative sample equalization function in distance learning, a weight policy of an asymmetric sample is also adopted to solve the asymmetry problem, where the weight policy of the asymmetric sample is as follows: once per iteration, num + + (i.e., num increases by 1), by weighting the parameters
Figure BDA0002697569290000143
And adjusting the weight value.
For the triple depth metric learning algorithm, in order to reduce the computation amount, only triplets which violate the constraint condition and reach a non-zero loss (i.e., non-zero loss) constraint condition are calculated.
At the same time, for those in formula (18)
Figure BDA0002697569290000144
The values are sorted from large to small, and sample matching is performed from the maximum value to reduce the interference of noise images.
After the vehicle re-identification scheme of the visual angle self-adaption and block position measurement learning is tested, the test result shows that: by adding visual angle transformation (pos transformation) and block position metric learning, the recognition rate is improved from 73.7% to 87.7% in the pairwise method, and the recognition rate is improved from 87.5% to 91.8% in the triplet method, so that the vehicle recognition rate is obviously improved.
Referring to fig. 3, a vehicle re-identification client is further provided as another embodiment of the present invention.
The client 200 includes a model training module 210 and a re-recognition module 220.
The model training module 210 is configured to input a vehicle image set including different vehicle images as input data into the convolutional neural network model for training until the model converges, so as to obtain a basic network model.
The re-identification module 220: acquiring a query vehicle input by a user, respectively acquiring a feature vector of each image in a query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by a distance measurement method, judging whether the query vehicle is similar to the image in the comparison library according to the distance, and outputting an identification result; the feature vector is a global feature generated by fusing a plurality of local block features fused into the view angle feature.
Other technical features refer to the foregoing embodiment, and the modules can be configured to execute the corresponding information processing method, which is not described herein again.
Referring to fig. 4, a vehicle re-identification system is further provided as another embodiment of the present invention.
The system 300 includes a user terminal 310 and a server 310.
The user terminal 310 is provided with a human-computer interaction interface, and a vehicle image set for training and query information input by a user are acquired through the human-computer interaction interface.
The server side 320 includes a processor and a memory for storing processor-executable instructions and parameters.
The processor is configured to:
inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model; and the number of the first and second groups,
acquiring a query vehicle input by a user, respectively acquiring a feature vector of each image in a query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by a distance measurement method, judging whether the query vehicle is similar to the image in the comparison library according to the distance, and outputting an identification result; the feature vector is a global feature generated by fusing a plurality of local block features fused into the view angle feature.
Other technical features refer to the previous embodiment, and the processor can be configured to execute the corresponding information processing method, which is not described herein again.
In the foregoing description, the disclosure of the present invention is not intended to limit itself to these aspects. Rather, the various components may be selectively and operatively combined in any number within the intended scope of the present disclosure. In addition, terms like "comprising," "including," and "having" should be interpreted as inclusive or open-ended, rather than exclusive or closed-ended, by default, unless explicitly defined to the contrary. All technical, scientific, or other terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs unless defined otherwise. Common terms found in dictionaries should not be interpreted too ideally or too realistically in the context of related art documents unless the present disclosure expressly limits them to that. Any changes and modifications of the present invention based on the above disclosure will be within the scope of the appended claims.

Claims (10)

1. A vehicle re-identification method is characterized by comprising a training process and a re-identification process, and the steps are as follows:
training process: inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model;
and (3) re-identification process: acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing a depth measurement learning network, implementing sequencing according to the distance, and outputting an identification result;
the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion.
2. The method of claim 1, wherein the training process is preceded by the step of constructing a set of vehicle images, as follows:
erecting a plurality of camera terminal groups in a real lane, wherein each camera terminal group comprises a plurality of camera terminal devices to form a group of linked camera groups of 0-180 degrees around a vehicle; each group of camera terminal groups comprises 7 camera terminal devices, and the visual angle is divided into 7 visual angles which are respectively 0 degree, 30 degrees, 60 degrees, 90 degrees, 120 degrees, 150 degrees and 180 degrees;
the method comprises the steps that a plurality of groups of vehicle video images and/or vehicle static images at a plurality of visual angles are collected through the camera terminal equipment; and screening the images, and selecting a clear picture corresponding to each view angle of each vehicle to be selected into a vehicle image set.
3. The method of claim 2, wherein: the training process comprises a feature extraction step, a measurement learning step and a classification step;
the characteristic extraction step comprises the steps of obtaining each image and extracting a plurality of local block characteristics of the vehicle in the image;
the metric learning step comprises the steps of converting the local block features into local block features fused with view angle features through a plurality of set basic feature views, carrying out fusion calculation on the local block features fused with the view angle features to obtain global features, and carrying out distance metric learning on the global features.
4. The method of claim 3, wherein: the local blocking features include a color histogram feature f of the host vehicle with a relative position dimensioncrFull pixel color histogram feature fchAnd texture feature fvThe feature extraction step is as follows:
vehicle image set with aspect ratio W × D for image set
Figure FDA0002697569280000011
Two different vehicles are taken from the training set as M and N,
Figure FDA0002697569280000012
the test vehicle is noted as T,
Figure FDA0002697569280000013
extracting a color component f of the host vehicle with a relative position dimension for each vehicle imagecrThe following are:
dividing the vehicle image into 12 x 12 non-overlapping blocks Pw，dAnd acquiring a block color histogram and recording the block color histogram as:
Figure FDA0002697569280000021
wherein K represents the number of dereferencing characteristics, and K represents the number of dereferencing characteristics; h (k) is the number of pixels in the image having a characteristic value of k; w and d are respectively the central horizontal and longitudinal axis coordinates of the block;
for dimension reduction and improvement of separability, the component values of the first 10% of h (k) in h (k) represent the block color characteristics and are recorded as
Figure FDA0002697569280000022
Meanwhile, the center position of the score block is (c)w，cd) The center position of the block distance graph is
Figure FDA0002697569280000023
Obtaining the block color characteristics containing the block position information as follows:
Figure FDA0002697569280000024
wherein pw and pd respectively represent the row and column numbers of the block, and fcrThe calculation formula of (a) is as follows:
Figure FDA0002697569280000025
where, s is 1, 2, …, pw × pd.
Based on formula (1), full-pixel color histogram information f is acquiredchThe following are:
Figure FDA0002697569280000026
obtaining texture features by extracting FV coding dense features as follows:
converting the vehicle image into a gray image, dividing the vehicle image into 24 × 24 non-overlapping blocks P for the vehicle images of multiple view anglesw′，d′Extracting low-level SIFT descriptors point by point from each 24 × 24 image block; w ', d ' are respectively the central horizontal and vertical axis coordinates of the block, and the central position of the block is marked as (c 'w，c′d)；
Dimensionality reduction by PCA algorithmGet SIFT descriptor from 128D
Figure FDA0002697569280000027
And adding the central point position to obtain the block color characteristics containing the block position information, as follows:
Figure FDA0002697569280000028
v to be extractedpatchPerforming feature fitting to obtain a Gaussian mixture model GMM, wherein the model comprises H Gaussian models, and enabling omega to be omegah，γh，ShRespectively representing the weight, mean and covariance matrix of the H model in the GMM, H is 0, 1, …, H-1; the first order differential v of each Gaussian model is calculated as follows(1)And a second order differential v(2)：
Figure FDA0002697569280000029
Figure FDA00026975692800000210
Wherein N represents the number of features;
Figure FDA0002697569280000031
representing the weight of the nth feature for the h gaussian model, N being 0, 1, …, N-1; x is the number ofnA feature value representing an nth feature;
will be provided with
Figure FDA0002697569280000032
And
Figure FDA0002697569280000033
are connected in series to obtain:
Figure FDA0002697569280000034
and adding position information to obtain (l +2) × H × 2 dimensional integral FV characteristics, wherein l is 128, and the FV characteristics are as follows:
Figure FDA0002697569280000035
sort (-) function sorting is carried out on the energy, Z key positions capable of representing the textural features of the energy are selected to form fvThe expression is as follows:
Figure FDA0002697569280000036
wherein ss is 1, 2, …, Z, Z is an integer greater than 1.
5. The method of claim 4, wherein: setting 0, 90 and 180 degrees as three basic feature perspectives, and calculating to obtain global features according to the following steps:
marking three basic characteristic visual angles i of 0 degree, 90 degrees and 180 degrees as V respectivelyfro、Vsid、VrevConverting all images in the training set into a picture characteristic function v blended with the visual angle characteristicsnullThe following are:
Figure FDA0002697569280000037
wherein alpha isi，θA view transformation characteristic function representing a view i;
the method for obtaining the characteristics of the blocks FV according to the formula (9) and the formula (11), the formula (9) can be expressed as the characteristics blended into the characteristics of the view angle:
Figure FDA0002697569280000038
is calculated to obtain
Figure FDA0002697569280000039
Simultaneously obtain f'ch，
Figure FDA00026975692800000310
Formula (3) is similarly rewritten and expressed as a feature that has been incorporated into the viewing angle feature as follows:
Figure FDA00026975692800000311
calculating to obtain a global feature FfullThe following are:
Ffull＝β1，θf′cr+β2，θf′ch+β3，θf′v (14)
wherein, betaj，θAnd j is more than or equal to 1 and less than or equal to 3, and is used as a classification characteristic category.
6. The method of claim 5, wherein: the distance measurement learning adopts the Mahalanobis distance, and for any two vehicle images, the step of calculating the distance function between the characteristic vectors of the two vehicle images through the Mahalanobis distance comprises the following steps:
for any two vehicle images si and sj, a distance function d (y) based on a characteristic vector between the images si and sj is calculated through a horse-type distancesi，ysj) The following are:
Figure FDA0002697569280000041
wherein, A is a diagonal matrix which represents the weight to distinguish the importance degree of the characteristic vector; decomposing A into WTW is solved instead of the weight beta in equation (14)j，θThe function of (1); z is a radical ofsiIs siCharacteristic vector of (2), zsiA feature vector of sj;
through rijE (1, -1) to express whether the two images are similar, and the similarity belongs to the same class, rij1, otherwise, r is not of the same classijThe constraint function is given by-1:
Figure FDA0002697569280000042
wherein the content of the first and second substances,
Figure FDA0002697569280000043
the critical value parameter represents the distance between classes and the class distance, delta (si, sj) represents a positive and negative sample equalization function, and the calculation formula of delta (si, sj) is as follows:
Figure FDA0002697569280000044
wherein z issnThe minimum value of the distance of all negative samples;
the classification step comprises: and (3) realizing the pairwise metric learning by calculating W, and obtaining final classification information by a KNN method.
7. The method of claim 6, wherein: in the training process, model training is carried out on a large image database by adopting a triple depth measurement learning algorithm, wherein the triple comprises a group of test images T and two groups of training images M and N; when the test acquisition test image T is close to M and far away from N, M, N is corresponding to the global feature { F } of Tfull(M)，Ffull(N)，Ffull(T) is adjusted to { y }M，yN，yTAnd (5) rewriting the distance constraint conditions of the feature vectors of the three components according to the formula (15) as follows:
Figure FDA0002697569280000045
wherein the content of the first and second substances,
Figure FDA0002697569280000053
is a critical value parameter of the distance between classes and the class distance; zT、ZN、ZMFeature vectors, z, for images T, M and N, respectivelysiIs the feature vector of sj.
8. The method of claim 7, wherein: in the training process, aiming at the pair wise method, the asymmetric problem is solved by adopting a weight strategy of an asymmetric sample, wherein the weight strategy of the asymmetric sample is as follows: once per iteration, num + +, by weight parameter
Figure FDA0002697569280000051
Adjusting the weight value; aiming at the triple depth metric learning algorithm, in order to reduce the operation weight, only triplets which violate the constraint condition and reach the non-zero loss constraint condition are calculated; and the number of the first and second groups,
for those in the formula (18)
Figure FDA0002697569280000052
The values are sorted from large to small, and sample matching is performed from the maximum value to reduce the interference of noise images.
9. A vehicle re-identification client, comprising the structure of:
the model training module is used for inputting a vehicle image set containing different vehicle images into the convolutional neural network model as input data to train until the model converges to obtain a basic network model;
and a re-identification module: the system comprises a base network model, a comparison library, a query vehicle image and a depth measurement learning network, wherein the base network model is used for acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and the comparison library based on the base network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing the depth measurement learning network, implementing sequencing according to the distance and outputting an identification result; the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion.
10. A vehicle re-identification system, characterized by: comprises a user terminal and a server terminal,
the user terminal is provided with a human-computer interaction interface, and a vehicle image set for training and query information input by a user are acquired through the human-computer interaction interface;
the server side includes a processor and a memory for storing processor-executable instructions and parameters, the processor configured to:
inputting a vehicle image set containing different vehicle images into a convolutional neural network model as input data to train until the model converges to obtain a basic network model; and the number of the first and second groups,
acquiring a query vehicle image input by a user, respectively acquiring a feature vector of each image in the query vehicle image and a comparison library based on the basic network model, calculating the distance between the feature vector of the query vehicle and the feature vector of the image in the comparison library by constructing a depth measurement learning network, implementing sequencing according to the distance, and outputting an identification result;
the feature vector is a global feature generated by adding a plurality of local block features blended into the view angle feature and the relative position information and then performing feature fusion.
CN202011011077.0A 2020-09-23 2020-09-23 Vehicle re-identification method, client and system Pending CN112418262A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011011077.0A CN112418262A (en) 2020-09-23 2020-09-23 Vehicle re-identification method, client and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011011077.0A CN112418262A (en) 2020-09-23 2020-09-23 Vehicle re-identification method, client and system

Publications (1)

Publication Number Publication Date
CN112418262A true CN112418262A (en) 2021-02-26

Family

ID=74854760

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011011077.0A Pending CN112418262A (en) 2020-09-23 2020-09-23 Vehicle re-identification method, client and system

Country Status (1)

Country Link
CN (1) CN112418262A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115761659A (en) * 2023-01-09 2023-03-07 南京隼眼电子科技有限公司 Recognition model construction method, vehicle type recognition method, electronic device, and storage medium

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463151A (en) * 2015-01-05 2015-03-25 北京邮电大学 Rapid target matching method based on fusion of main color histograms and spatial position information
CN108197538A (en) * 2017-12-21 2018-06-22 浙江银江研究院有限公司 A kind of bayonet vehicle searching system and method based on local feature and deep learning
US20180178792A1 (en) * 2016-12-26 2018-06-28 Toyota Jidosha Kabushiki Kaisha Vehicle identification device and vehicle identification method
CN109583305A (en) * 2018-10-30 2019-04-05 南昌大学 A kind of advanced method that the vehicle based on critical component identification and fine grit classification identifies again
CN109740479A (en) * 2018-12-25 2019-05-10 苏州科达科技股份有限公司 A kind of vehicle recognition methods, device, equipment and readable storage medium storing program for executing again
CN110309770A (en) * 2019-06-28 2019-10-08 华侨大学 A kind of vehicle discrimination method again based on the study of four-tuple loss metric
CN110516640A (en) * 2019-08-30 2019-11-29 华侨大学 It is a kind of to combine the vehicle indicated discrimination method again based on feature pyramid
CN111291821A (en) * 2020-02-20 2020-06-16 上海眼控科技股份有限公司 Vehicle weight recognition method and device, computer equipment and storage medium

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104463151A (en) * 2015-01-05 2015-03-25 北京邮电大学 Rapid target matching method based on fusion of main color histograms and spatial position information
US20180178792A1 (en) * 2016-12-26 2018-06-28 Toyota Jidosha Kabushiki Kaisha Vehicle identification device and vehicle identification method
CN108197538A (en) * 2017-12-21 2018-06-22 浙江银江研究院有限公司 A kind of bayonet vehicle searching system and method based on local feature and deep learning
CN109583305A (en) * 2018-10-30 2019-04-05 南昌大学 A kind of advanced method that the vehicle based on critical component identification and fine grit classification identifies again
CN109740479A (en) * 2018-12-25 2019-05-10 苏州科达科技股份有限公司 A kind of vehicle recognition methods, device, equipment and readable storage medium storing program for executing again
CN110309770A (en) * 2019-06-28 2019-10-08 华侨大学 A kind of vehicle discrimination method again based on the study of four-tuple loss metric
CN110516640A (en) * 2019-08-30 2019-11-29 华侨大学 It is a kind of to combine the vehicle indicated discrimination method again based on feature pyramid
CN111291821A (en) * 2020-02-20 2020-06-16 上海眼控科技股份有限公司 Vehicle weight recognition method and device, computer equipment and storage medium

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
YI ZHOU ET AL.: "Vehicle Re-Identification by Deep Hidden Multi-View Inference", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 *
王茜: "基于监控的大型城市车辆目标识别与分类的若干关键技术研究", 《中国博士学位论文全文数据库 信息科技辑》 *
蔡晓东 等: "基于多分支卷积神经网络的车辆图像比对方法", 《电视技术》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115761659A (en) * 2023-01-09 2023-03-07 南京隼眼电子科技有限公司 Recognition model construction method, vehicle type recognition method, electronic device, and storage medium

Similar Documents

Publication Publication Date Title
CN107609601B (en) Ship target identification method based on multilayer convolutional neural network
CN108573276B (en) Change detection method based on high-resolution remote sensing image
CN113065558A (en) Lightweight small target detection method combined with attention mechanism
CN113160192B (en) Visual sense-based snow pressing vehicle appearance defect detection method and device under complex background
Chen et al. Vehicle detection in high-resolution aerial images via sparse representation and superpixels
Zhou et al. Robust vehicle detection in aerial images using bag-of-words and orientation aware scanning
CN109558823B (en) Vehicle identification method and system for searching images by images
CN107633226B (en) Human body motion tracking feature processing method
CN112633382B (en) Method and system for classifying few sample images based on mutual neighbor
CN112686812B (en) Bank card inclination correction detection method and device, readable storage medium and terminal
CN109063649B (en) Pedestrian re-identification method based on twin pedestrian alignment residual error network
US20090060340A1 (en) Method And Apparatus For Automatic Image Categorization Using Image Texture
Zhang et al. Road recognition from remote sensing imagery using incremental learning
CN111709313B (en) Pedestrian re-identification method based on local and channel combination characteristics
CN112150493A (en) Semantic guidance-based screen area detection method in natural scene
CN108932518B (en) Shoe print image feature extraction and retrieval method based on visual bag-of-words model
CN111046789A (en) Pedestrian re-identification method
CN113592923B (en) Batch image registration method based on depth local feature matching
CN110929746A (en) Electronic file title positioning, extracting and classifying method based on deep neural network
CN109635726B (en) Landslide identification method based on combination of symmetric deep network and multi-scale pooling
CN109325407B (en) Optical remote sensing video target detection method based on F-SSD network filtering
CN110738100A (en) camouflage military target identification method and system based on deep learning
CN112818905A (en) Finite pixel vehicle target detection method based on attention and spatio-temporal information
CN114596500A (en) Remote sensing image semantic segmentation method based on channel-space attention and DeeplabV3plus
CN116385958A (en) Edge intelligent detection method for power grid inspection and monitoring

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210226

RJ01 Rejection of invention patent application after publication