CN111401113A - Pedestrian re-identification method based on human body posture estimation - Google Patents

Pedestrian re-identification method based on human body posture estimation Download PDF

Info

Publication number
CN111401113A
CN111401113A CN201910006045.2A CN201910006045A CN111401113A CN 111401113 A CN111401113 A CN 111401113A CN 201910006045 A CN201910006045 A CN 201910006045A CN 111401113 A CN111401113 A CN 111401113A
Authority
CN
China
Prior art keywords
picture
pedestrian
human body
neural network
network
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
CN201910006045.2A
Other languages
Chinese (zh)
Inventor
于耀
邱睿
周余
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.)
Nanjing University
Original Assignee
Nanjing University
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 Nanjing University filed Critical Nanjing University
Priority to CN201910006045.2A priority Critical patent/CN111401113A/en
Publication of CN111401113A publication Critical patent/CN111401113A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • 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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a pedestrian re-recognition algorithm based on human body posture estimation, and belongs to the field of computer vision. The invention solves the problems that: the pedestrian re-identification task has the problem that the identification effect is not robust due to the fact that the human body posture and the camera angle are not fixed. The key part of the main algorithm of the invention is that human body joint points in a pedestrian picture are obtained by utilizing a human body posture estimation algorithm, and the pedestrian picture is divided into pictures of all parts of the human body through the joint points. The neural network is trained by pictures of each part of a human body, and the feature vectors of the parts are extracted. And further, fusing the feature vectors, and finally, measuring the distance of the feature vectors. The method and the device can improve the robustness of the unfixed sample of the pedestrian posture in the task of re-identifying the pedestrian, and have certain innovativeness.

Description

Pedestrian re-identification method based on human body posture estimation
Technical Field
The invention belongs to the field of computer vision and deep learning, and mainly relates to a pedestrian re-identification method based on attitude estimation.
Background
Pedestrian re-identification is a key task in a multi-camera monitoring system. The task requires that whether pedestrians shot by two different cameras are the same person or not is identified at different time and different places. Pedestrian re-identification is a key technology for realizing cross-camera pedestrian tracking and key person retrieval. Related research is increasing, but still there are many difficulties to be solved. Such as different resolution in the pictures taken by different cameras, illumination variations, irrelevant background noise. Meanwhile, the human body is not rigid, the actions of pedestrians are different, the shooting angles of the cameras are different, and the problem of posture change is also introduced. The above problems all affect the accuracy of the recognition to a different extent.
The existing pedestrian re-identification framework mainly comprises feature extraction and feature measurement. However, most methods ignore the influence of pedestrian posture change in the process of extracting image features, or perform implicit human body part alignment operation by training a neural network. Such an operation loses the spatial relationship of partial images of a part of the human body, so that the accuracy is reduced on matching some images with large differences in human body postures.
Disclosure of Invention
The invention provides a pedestrian re-identification method based on human body posture estimation. The method comprises the following steps:
step 1: the method comprises the steps of taking frames of pedestrian videos collected by a plurality of cameras, and detecting pedestrians in each frame of image;
step 2: acquiring the joint point position of a pedestrian in the picture through attitude estimation, and accordingly dividing the pedestrian picture into an upper body picture and a lower body picture;
and step 3: training 3 neural networks, and respectively extracting feature vectors of the original picture, the upper half body picture and the lower half body picture;
and 4, step 4: feature fusion and matching sorting.
For step 1, a plurality of cameras are distributed on different streets to shoot different scenes. And taking frames of the recorded video, wherein each frame of picture comprises pedestrians with different numbers. And obtaining a bounding box of each pedestrian by adopting a pedestrian detection method based on a convolutional neural network. And (4) carrying out screenshot on the bounding box of each pedestrian to obtain a pedestrian foreground picture.
And 2, performing joint point prediction on the pedestrian in the foreground picture by adopting a method based on a convolutional neural network. All joint points are two-dimensional joint points, which are labeled in the form of (x, y), with x representing the joint point abscissa and y representing the ordinate. By the ordinate y of the left and right hip joints1、y2Calculate the average value
Figure BSA0000176770920000011
Figure BSA0000176770920000012
To be provided with
Figure BSA0000176770920000013
As a boundary, the pedestrian picture is divided into an upper half and a lower half. The human body is a non-rigid object, and various postures can appear in a pedestrian picture shot by a street monitoring camera. Compared with the division of other image characteristic methods, the division of each part of the human body based on the joint points has better resistance to adverse effects caused by human body shape change. The overall deformation of the upper body and the lower body is relatively small; meanwhile, the part contains large picture areas and cannot be completely blocked.
And 3, respectively forming three independent training sets by the pedestrian whole-body picture, the upper-half body picture and the lower-half body picture obtained by posture estimation. And the same strategy is adopted to train three convolutional neural networks with the same structure and independent weights. And preprocessing the upper half body picture, the lower half body picture and the original picture. The preprocessing step is to zoom the picture to make the whole picture be square; and then randomly turning the picture left and right. And inputting the upper and lower half-body pictures after the preprocessing together with the original pictures into a neural network for two-stage training. The first stage is training with a Softmax loss function, and the formula is as follows:
Figure BSA0000176770920000021
wherein a isjIs the j-th type of linear prediction score, y, of the network outputiIs the label of the input sample corresponding to class i, SiRepresenting the probability that the current sample corresponds to category i, and T representing the total number of types L1Is a numerical result of the loss function. When the neural network is trained by using the Softmax loss function to reach a certain precision, training is carried outAnd entering the second stage, and continuing by using the triple loss function. The Triplet loss function equation is as follows:
Figure BSA0000176770920000023
where f (x) represents the mapping from the input picture to the neural network feature vector.
Figure BSA0000176770920000024
Is a pedestrian picture randomly selected from the data set,
Figure BSA0000176770920000025
represents another picture with the same label,
Figure BSA0000176770920000026
a picture representing a different label.
Figure BSA0000176770920000027
The total number of all triples in the training set is n. α, which is the margin value, i.e., the boundary padding value.
Figure BSA0000176770920000028
One term obtains the feature vector f (x)a)、f(xp) The square of the two norms between them, namely the squared Euclidean distance, represents the intra-class distance of the feature vectors of the same type of samples. In the same way
Figure BSA0000176770920000029
One term characterizes the inter-class distance between the heterogeneous sample feature vectors. Only when f (x)a) And f (x)p) Squared Euclidean distance between, and f (x)a) And f (x)n) When the squared Euclidean distance difference is greater than α, the loss function value is greater than zero, and the optimization of the network is guidedAnd aggregating the features of the same pedestrian picture and dispersing the features of different pedestrian pictures.
The method is characterized in that when a Triplet loss function is used for training a neural network, a training sample picture needs to be divided into triplets serving as input data, a large number of Triplet combinations can obtain zero values in the calculation of the Triplet loss function, the zero values do not contribute to network training, and the triplets are excluded, so that a difficult example mining method based on a small batch gradient Descent method is adopted in the network training processa＝[x1，x2...xn]Wherein x isiRepresenting the ith input sample after sorting, a unit coefficient c is set, and L ist is pairedaRandomly scrambling the groups, scrambling the order between groups, and leaving the order within the groups unchanged to obtain L istbFinally set mini batch size, L istbA slicing operation is performed, whereby all the mini batch required for training is obtained.
Each training iteration only uses a sample in a mini batch, and a triple is constructed in each training by adopting the following method that all similar samples in the mini batch are combined, all samples which are different from the samples in the mini batch are found out for each similar sample and are combined together to obtain a triple, all the obtained triplets in the mini batch are input into a network, and the final triple loss function value L is calculated2All triplets are as per L2The sorting is done from large to small. The top ranked samples are called difficult cases. But such asIf the network is trained with all use difficulties, the network training will likely collapse. A hard-case coefficient λ between 0 and 1 is set to represent the proportion of hard-case triples used in one training iteration to the total number. And setting a coefficient n to indicate that the number of the triples entering the training in one iteration is n. Then the last triplet set participating in training is ranked by
Figure BSA0000176770920000032
A triplet difficulty, and
Figure BSA0000176770920000033
a randomly selected triplet.
And inputting a whole body picture, an upper body picture and a lower body picture with a pedestrian picture query into the corresponding network respectively by taking the trained network as a feature extractor. The output of the last fully connected layer of the network is extracted as a feature vector.
For step 4, connecting the three part of feature vectors obtained in step 3 end to form a combined feature vector, wherein the formula is as follows:
F′＝[F1F2F3]
wherein F1、F2、F3Respectively representing feature vectors extracted from the pictures of the whole body, the upper half and the lower half. The combined feature vector retains the whole information of the picture and the local information of the upper and lower bodies of the pedestrian, and has better robustness. Meanwhile, the integral features and the local features of the pedestrians are fixed on the specific vector dimension in a combined feature vector mode, namely, the integral features and the local features are subjected to explicit alignment operation, and the influence of the attitude change of the pedestrians on the accuracy rate of the heavy recognition can be reduced.
And finally, performing characteristic distance measurement. The existing candidate pedestrian picture set G to be inquired needs to find out the target pedestrian in the searched picture p. Extracting a joint feature vector F 'of the picture p'pAnd a joint feature vector set G of all pictures in the picture set. And calculating Euclidean distances between p and all the joint feature vectors in G.
G＝{x1，x2，...，xm}
G′＝{F′1，F′2，...，F′m}
Figure BSA0000176770920000031
Wherein F'iRepresenting the joint feature vector of the ith sample within the set G. diIs represented by F'pAnd F'iEuclidean distance of. diThe smaller the size, the more similar the pedestrian in the sample is to the target pedestrian. Sample x in GiAccording to diThe values of (c) are ordered from small to large. And outputting the candidate pictures with the fixed number of the top ranks.
Drawings
The accompanying drawings are included to provide a further understanding of the invention, and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. The drawings illustrate the following:
fig. 1 is a system architecture diagram.
Fig. 2 is a distribution diagram of the joint points.
FIG. 3 is a combined feature extraction graph.
Detailed Description
The embodiments of the present invention will be described in detail with reference to the accompanying drawings, by which how to apply the technical means to solve the problems and achieve the technical effects can be fully understood and implemented. The preferred examples are given solely for the purpose of illustration and are not intended to limit the scope of the invention.
The following describes the execution of the algorithm in detail. As shown in fig. 1, the pedestrian re-identification method based on human body posture estimation of the present invention comprises the following steps:
step 1: multi-camera pedestrian detection
Cameras are respectively installed on different streets. And taking frames of the recorded video, wherein each frame of picture comprises pedestrians with different numbers. And obtaining a foreground picture of each pedestrian by adopting a pedestrian detection method based on a convolutional neural network.
Step 2: pedestrian attitude estimation
And predicting the joint points by using a pedestrian posture estimation method based on a neural network. As shown in fig. 2, the pedestrian posture estimation method outputs the positions of the pixels of the respective joint points in the image. These joint points are nose, neck, left eye, right eye, left ear, right ear, left shoulder, right shoulder, left elbow, right elbow, left wrist, right wrist, left hip, right hip, left knee, right knee, left ankle, and right ankle, respectively. Each joint point has a dimension of 2, which is of the form (x, y). x is the abscissa of the pixel where the joint point is located, and y is the ordinate. By the ordinate value y of the left and right hip joints1、y2Calculate the average value
Figure BSA0000176770920000041
Figure BSA0000176770920000042
Finally, to
Figure BSA0000176770920000043
As a boundary, the pedestrian foreground picture is divided into an upper-body picture and a lower-body picture.
And step 3: training neural network to extract feature vector
To better train the neural network, the upper body picture, the lower body picture and the whole body picture obtained in step 2 are first preprocessed, the pictures are scaled so that the whole picture is square in size, preferably set to a picture of 256 ＊ 256 pixels, then they are randomly mirror flipped left and right, preferably the random ratio of the mirror flipping left and right of the pictures is set to 50%. the average of all picture pixels R, G, B three channels is calculated, set to be 50%. fig
Figure BSA0000176770920000044
And subtracting the mean value of the corresponding channel from all the picture pixels to realize mean value regularization.
Two stages of training are used in the neural network training process. The two stages differ in the loss function used to train the network. The first stage adopts the first stage to train by adopting a Softmax loss function, and the formula is as follows:
Figure BSA0000176770920000045
Figure BSA0000176770920000051
wherein a isiIs the i-th type of linear prediction score, y, of the network outputiIs a label of the input sample, SiIs the probability that the current sample corresponds to the ith type L1Is the last calculated loss function value, in the first phase, when the network gradually converges, finally L1And entering the second stage when the value is not reduced any more. Training is continued with the triple loss function instead. The Triplet loss function is formulated as follows:
Figure BSA0000176770920000052
where f (x) represents the mapping from the input picture to the neural network feature vector.
Figure BSA0000176770920000053
Is a pedestrian picture randomly selected from the data set,
Figure BSA0000176770920000054
represents another picture with the same label,
Figure BSA0000176770920000055
a picture representing any one of the different labels.
Figure BSA0000176770920000056
All triplets in the training set are n. α, the total number of triplets is margin, i.e. boundary padding value, preferably α is set to 0.5.
During the second stage of training, a difficult case mining method based on a small batch gradient descent method is adopted, the input pictures are sorted according to the labels, and a sequence L ist is obtaineda＝[x1，x2...xn]Wherein x isiRepresenting the ith input sample after sorting, a unit coefficient c is set, preferably 4, every 4 samples form a sample subgroup, and L ist is paired with the subgroup as the minimum unitaPerforming a scrambling operation to obtain L istbPreferably the size of the sample in the mini batch is set to 40. with 40 as the step size L ist will be usedbA slicing operation is performed, whereby all the mini batch required for training is obtained. After the mini batch is obtained, the sample is dug difficultly. A difficult-to-excavate scaling factor λ is set, preferably λ is set to 0.2 at the start of training and λ is set to 0.5 in the middle of training. And setting a coefficient n to indicate that the number of the triples entering the training in one iteration is n. Then the last triplet set participating in training is ranked by
Figure BSA0000176770920000057
A triplet difficulty, and
Figure BSA0000176770920000058
a randomly selected triplet.
As shown in fig. 3, the upper body picture, the lower body picture and the whole body picture are trained by the above steps to obtain a network 1, a network 2 and a network 3. And extracting the output of the last-but-one layer of the fully connected layer of the network, preferably setting the number of neurons of the layer to be 256, and setting the output of the layer to be a feature vector.
And 4, step 4: feature fusion and matching
And inputting the whole body picture, the upper half body picture and the lower half body picture of the picture to be inquired into the neural network. And splicing the obtained feature vectors end to form a 768-dimensional combined feature vector. And calculating the Euclidean distance between the combined feature vector of the target pedestrian picture and the combined feature vector of the pedestrian picture set to be retrieved. And sequencing the pedestrian picture sets to be retrieved from small to large according to the Euclidean distance. Preferably, the first 10 pictures of the pedestrian to be retrieved are output as a result.
Although the embodiments of the present invention have been shown and described, the above descriptions are only for the convenience of understanding the present invention and are not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A pedestrian re-identification method based on human body posture estimation is characterized by comprising the following main steps:
step one, taking frames of pedestrian videos collected by a plurality of cameras, detecting pedestrians in each frame of image, and obtaining a pedestrian foreground picture;
acquiring the joint point position of the pedestrian in the pedestrian picture by using a posture estimation method, and accordingly dividing the pedestrian picture into an upper body picture and a lower body picture;
step three, training a neural network, extracting features of the original picture, the upper half body picture and the lower half body picture through the neural network, and outputting feature vectors of all parts of the pedestrian;
and step four, fusing the obtained feature vectors of all the parts to obtain a combined feature vector. And (5) carrying out distance measurement through the joint feature vector, and outputting a final result.
2. The second step of the method of claim 1 is characterized in that joint prediction is performed on the pedestrian foreground picture by using a human body posture estimation method based on a convolutional neural network. And (3) marking the pixel position of the two-dimensional joint point obtained according to the prediction by (x, y), wherein x represents the abscissa of the joint point pixel, and y represents the ordinate of the joint point pixel. By the ordinate y of the left and right hip joints1，y2Calculate the average value
Figure FSA0000176770910000011
Figure FSA0000176770910000012
To be provided with
Figure FSA0000176770910000013
As a boundary, the pedestrian picture is divided into an upper body picture and a lower body picture.
3. Step three of the method of claim 1 wherein the neural network is trained in two stages using different loss functions. The first stage uses a Softmax loss function, which is formulated as follows:
Figure FSA0000176770910000014
Figure FSA0000176770910000015
aiis the i-th type of linear prediction score, y, of the network outputiIs a label of the input sample, SiIs the probability that the current sample corresponds to the ith type. When the network converges and the loss function value does not drop any more, the triple loss function is used, and the formula is as follows:
Figure FSA0000176770910000016
f (x) is the mapping of the input picture to the feature vector.
Figure FSA0000176770910000017
Is a randomly selected picture of a pedestrian,
Figure FSA0000176770910000018
is the same picture of another label,
Figure FSA0000176770910000019
a picture of a different label.
Figure FSA00001767709100000110
The total number of tuples in the training set is N, and α is the boundary padding value.
4. Step three of the method of claim 1 is characterized in that the method of claim 3 is adopted for the pedestrian's whole body picture, upper body picture and lower body picture, and the neural network is trained independently. And extracting the output of the last full-connection layer of the network to obtain the overall characteristic expression and the local characteristic expression of each sample.
5. A fourth step of the method of claim 1 is characterized in that the three part feature expressions obtained in the third step are vector-spliced to obtain a joint feature vector, and the global feature information and the local feature information of different samples are explicitly fixed in the same vector dimension.
CN201910006045.2A 2019-01-02 2019-01-02 Pedestrian re-identification method based on human body posture estimation Pending CN111401113A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910006045.2A CN111401113A (en) 2019-01-02 2019-01-02 Pedestrian re-identification method based on human body posture estimation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910006045.2A CN111401113A (en) 2019-01-02 2019-01-02 Pedestrian re-identification method based on human body posture estimation

Publications (1)

Publication Number Publication Date
CN111401113A true CN111401113A (en) 2020-07-10

Family

ID=71432230

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910006045.2A Pending CN111401113A (en) 2019-01-02 2019-01-02 Pedestrian re-identification method based on human body posture estimation

Country Status (1)

Country Link
CN (1) CN111401113A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112990144A (en) * 2021-04-30 2021-06-18 德鲁动力科技(成都)有限公司 Data enhancement method and system for pedestrian re-identification
CN113255695A (en) * 2021-05-21 2021-08-13 广州广电运通金融电子股份有限公司 Feature extraction method and system for target re-identification
CN113378729A (en) * 2021-06-16 2021-09-10 西安理工大学 Pose embedding-based multi-scale convolution feature fusion pedestrian re-identification method
CN114663917A (en) * 2022-03-14 2022-06-24 清华大学 Multi-view-angle-based multi-person three-dimensional human body pose estimation method and device
CN116052220A (en) * 2023-02-07 2023-05-02 北京多维视通技术有限公司 Pedestrian re-identification method, device, equipment and medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106650827A (en) * 2016-12-30 2017-05-10 南京大学 Human body posture estimation method and system based on structure guidance deep learning
CN106778527A (en) * 2016-11-28 2017-05-31 中通服公众信息产业股份有限公司 A kind of improved neutral net pedestrian recognition methods again based on triple losses
CN107832672A (en) * 2017-10-12 2018-03-23 北京航空航天大学 A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778527A (en) * 2016-11-28 2017-05-31 中通服公众信息产业股份有限公司 A kind of improved neutral net pedestrian recognition methods again based on triple losses
CN106650827A (en) * 2016-12-30 2017-05-10 南京大学 Human body posture estimation method and system based on structure guidance deep learning
CN107832672A (en) * 2017-10-12 2018-03-23 北京航空航天大学 A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112990144A (en) * 2021-04-30 2021-06-18 德鲁动力科技(成都)有限公司 Data enhancement method and system for pedestrian re-identification
CN113255695A (en) * 2021-05-21 2021-08-13 广州广电运通金融电子股份有限公司 Feature extraction method and system for target re-identification
CN113378729A (en) * 2021-06-16 2021-09-10 西安理工大学 Pose embedding-based multi-scale convolution feature fusion pedestrian re-identification method
CN113378729B (en) * 2021-06-16 2024-05-10 西安理工大学 Multi-scale convolution feature fusion pedestrian re-identification method based on pose embedding
CN114663917A (en) * 2022-03-14 2022-06-24 清华大学 Multi-view-angle-based multi-person three-dimensional human body pose estimation method and device
CN116052220A (en) * 2023-02-07 2023-05-02 北京多维视通技术有限公司 Pedestrian re-identification method, device, equipment and medium
CN116052220B (en) * 2023-02-07 2023-11-24 北京多维视通技术有限公司 Pedestrian re-identification method, device, equipment and medium

Similar Documents

Publication Publication Date Title
CN107832672B (en) Pedestrian re-identification method for designing multi-loss function by utilizing attitude information
CN111160297B (en) Pedestrian re-identification method and device based on residual attention mechanism space-time combined model
CN111259850B (en) Pedestrian re-identification method integrating random batch mask and multi-scale representation learning
Zhong et al. Grayscale enhancement colorization network for visible-infrared person re-identification
Xie et al. Comparator networks
CN112861720B (en) Remote sensing image small sample target detection method based on prototype convolutional neural network
Laskar et al. Camera relocalization by computing pairwise relative poses using convolutional neural network
Bhagat et al. Indian sign language gesture recognition using image processing and deep learning
CN111401113A (en) Pedestrian re-identification method based on human body posture estimation
CN111325115B (en) Cross-modal countervailing pedestrian re-identification method and system with triple constraint loss
CN111814661B (en) Human body behavior recognition method based on residual error-circulating neural network
CN107145826B (en) Pedestrian re-identification method based on double-constraint metric learning and sample reordering
CN109800794B (en) Cross-camera re-identification fusion method and system for appearance similar targets
CN109472191B (en) Pedestrian re-identification and tracking method based on space-time context
CN110717411A (en) Pedestrian re-identification method based on deep layer feature fusion
CN111709311A (en) Pedestrian re-identification method based on multi-scale convolution feature fusion
CN110909605A (en) Cross-modal pedestrian re-identification method based on contrast correlation
CN111738048B (en) Pedestrian re-identification method
CN112001278A (en) Crowd counting model based on structured knowledge distillation and method thereof
CN112507853B (en) Cross-modal pedestrian re-recognition method based on mutual attention mechanism
CN113610046B (en) Behavior recognition method based on depth video linkage characteristics
TW202141424A (en) Target tracking method and apparatus, storage medium
CN111639580A (en) Gait recognition method combining feature separation model and visual angle conversion model
CN111582154A (en) Pedestrian re-identification method based on multitask skeleton posture division component
CN112329662B (en) Multi-view saliency estimation method based on unsupervised learning

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200710

WD01 Invention patent application deemed withdrawn after publication