CN109961051B - Pedestrian re-identification method based on clustering and block feature extraction - Google Patents
Pedestrian re-identification method based on clustering and block feature extraction Download PDFInfo
- Publication number
- CN109961051B CN109961051B CN201910243050.5A CN201910243050A CN109961051B CN 109961051 B CN109961051 B CN 109961051B CN 201910243050 A CN201910243050 A CN 201910243050A CN 109961051 B CN109961051 B CN 109961051B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- clustering
- loss function
- feature extraction
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
-
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a pedestrian re-identification method based on clustering and block feature extraction, which comprises the following steps of (1) clustering images through K-means, respectively inputting the clustered images into a DCGAN network, respectively generating images, and expanding an original training set; (2) In deep learning, feature extraction is carried out on real data and generated data without labels through block feature extraction, meanwhile, the clustering labeling smooth normalization loss function (CLS) is adopted to carry out labeling training on the data, and rearrangement (Re-ranking) is adopted during testing to further improve the effect of pedestrian Re-identification. The invention combines the clustering labeling smooth normalization loss function and the block feature extraction method, solves the problems of limited pedestrian re-identification training data and label distribution, and simultaneously extracts the effective features of the image in a block mode.
Description
Technical Field
The invention belongs to the field of digital image processing and computer vision, relates to a pedestrian re-identification method, and particularly relates to a pedestrian re-identification method based on a generated countermeasure network and a convolutional neural network.
Background
Pedestrian re-identification is a technology for judging whether a specific pedestrian exists in an image or a video sequence by using a computer vision technology, and is widely considered as a sub-problem of image retrieval, namely, given a monitored pedestrian image, the pedestrian image under the cross-device is retrieved. The pedestrian re-identification has wide application prospects including pedestrian retrieval, pedestrian tracking, street event detection, pedestrian action and behavior analysis and the like.
In the field of computer vision, the object of pedestrian re-identification is to specify a pedestrian image, and identify the image of the pedestrian from the image library of the pedestrian under the visual angle of other existing non-overlapped cameras. In surveillance video, very high quality face pictures are often not available due to camera resolution and shooting angle. When face recognition fails, pedestrian re-recognition becomes a very important substitute technology. With the development of deep learning in recent years, pedestrian re-identification has taken a very great breakthrough. Deep learning has become a focus of research in the field of computer vision. Because the convolutional neural network has the capability of learning how to extract features, the convolutional neural network is more suitable for practical application in the engineering field than the traditional method.
Because a depth model needs a large amount of training data, but the acquisition of pedestrian re-identification data sets needs to manually perform target framing and ID calibration, which is a relatively high-cost data acquisition, a method for rapidly generating more pedestrian re-identification training data through GAN becomes a popular research direction.
Pedestrian re-identification is a difficult problem, and many challenges are faced in solving the problem. These challenges can be summarized in two categories: the first challenge is the need for large amounts of training data; the second challenge is a non-ideal scenario.
The pedestrian re-identification research mainly comprises the following steps: a feature representation based method and a method for generating a countermeasure network. The characteristic representation method is mainly used for researching and extracting identification characteristics with robustness to represent pedestrians, and the other way for improving the performance of the pedestrian re-identification model is to generate a new pedestrian image by means of a generation countermeasure network (GAN) so as to expand training data.
At present, the following defects mainly exist in the pedestrian re-identification:
(1) Limited training data;
from the current situation of collecting pedestrian re-recognition training data, the spatiotemporal distribution of the collected data relative to the real data is very limited and local. Meanwhile, the data size for pedestrian re-identification is also very small compared to other vision tasks. For example, in the case of the large-scale image recognition data set ImageNet, the training data of the large-scale image recognition data set ImageNet comprises 125 ten thousand pictures, and the pedestrian re-recognition data set currently frequently used for pedestrian re-recognition comprises only 3 ten thousand pictures of pedestrians.
The pedestrian re-identification training and data acquisition are difficult, and it is difficult for people to collect pedestrian data of cross time, cross climate and multiple scenes. In addition, privacy concerns also hinder data acquisition.
(2) The pedestrian re-identification data is difficult to label;
firstly, the labeling workload is huge, and the labeling cost is very large in terms of time and money. Secondly, the annotation itself is sometimes very difficult, and it is difficult to separate two people of similar age and appearance from each other in a video, wearing the same clothing.
(3) Non-ideal scenes;
the pedestrian identification method mainly comprises the steps that pedestrians are in different postures, complex backgrounds, different lighting conditions and different shooting visual angles exist, and great troubles are brought to the pedestrian re-identification. The pedestrian image has the problems of pedestrian misalignment, partial occlusion, low image quality and the like.
Disclosure of Invention
In order to solve the technical problem, the invention provides a pedestrian re-identification method based on generation of a countermeasure network and a convolutional neural network.
The technical scheme adopted by the invention is as follows: a pedestrian re-identification method based on clustering and block feature extraction is characterized by comprising the following steps:
step 1: collecting Pedestrian images under a monitoring camera to obtain a Pedestrian image library Pedestrian01, and carrying out k-means clustering on the Pedestrian 01;
step 2: respectively inputting the clustered images into a DCGAN network, and respectively generating unlabeled images;
and step 3: performing label distribution on the label-free image generated in the step 2 by adopting a clustering label smooth normalized loss function (CLS) to obtain a labeled generated image;
and 4, step 4: fusing the collected Pedestrian image library with the generated image obtained in the step 3, and expanding the collected Pedestrian image library to obtain a new Pedestrian image library Pedestrian02;
and 5: dividing the Peer 02 into p blocks horizontally, and inputting each block into a CNN network for feature extraction to obtain local features of the image;
step 6: performing combined training on the CNN network by adopting a clustering label smooth normalized loss function CLS and a cross entropy loss function in the training process;
and 7: during testing, re-ranking Re-ranking optimization is adopted, and a pedestrian Re-identification result is output.
Compared with the existing algorithm, the method has the remarkable advantages that:
(1) Aiming at the problems of difficult pedestrian re-identification training and data acquisition, the invention adopts a DCGAN network to expand a data set. Firstly, clustering images of an original data set by adopting a K-means method, dividing the images with similar characteristics into the same class, and then respectively sending the clustered K-class images into a DCGAN network to train so as to obtain K-class generated images. The clustering process makes the generated image more real and effective.
(2) Compared with the label smooth normalization, the cluster smooth normalization loss function (CLS) provided by the invention has more adaptability to the label of the generated sample because the class labels in different clusters are removed, and the classes in the same cluster are distributed with uniform probability, thereby avoiding over-concentration on a certain class and solving the problems of label distribution and over-smoothing.
(3) The invention mainly solves the problem of non-ideal scenes by detecting and matching human body parts and adopting a block feature extraction method, so that a network can learn more potential factors and the robustness is enhanced.
(4) The pedestrian re-identification method based on clustering and block feature extraction can be applied to re-identification of pedestrians in complex scenes, has the advantages of strong transportability, stable algorithm and high speed for scene change, can effectively solve the problems of small data set and label distribution, and has strong practicability.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Detailed Description
In order to facilitate the understanding and implementation of the present invention for those of ordinary skill in the art, the present invention is further described in detail with reference to the accompanying drawings and examples, it is to be understood that the embodiments described herein are merely illustrative and explanatory of the present invention and are not restrictive thereof.
The method of the invention is divided into two parts: (1) Clustering the images through K-means, respectively inputting the clustered images into a DCGAN network, respectively generating images, and expanding an original training set; (2) In deep learning, feature extraction is carried out on real data and generated data without labels through block feature extraction, meanwhile, the clustering labeling smooth normalization loss function (CLS) is adopted to carry out labeling training on the data, and rearrangement (Re-ranking) is adopted during testing to further improve the effect of Re-identification of the pedestrians. The method combines the clustering labeling smooth normalization loss function and the block feature extraction method, solves the problems of limited pedestrian re-identification training data and label distribution, and extracts the effective features of the image in a block mode.
Referring to fig. 1, the pedestrian re-identification method based on clustering and block feature extraction provided by the invention comprises the following steps:
step 1: collecting Pedestrian images under a monitoring camera to obtain a Pedestrian image library Peerstrian 01, and carrying out k-means clustering on the Peerstrian 01;
the specific implementation of the step 1 comprises the following substeps:
step 1.1: pedestrian images Pedestrian01 to ResNet50 network input to the monitoring camera adopt a cross entropy loss functionTraining to obtain a feature extraction model; wherein y is the original actual output,is the desired output;
step 1.2: inputting the training data set into the feature extraction model of step 1.1, and extracting the feature mapping vector x of the last convolutional layer n ;
Step 1.3: randomly selecting K feature mapping objects, wherein each object represents an initial mean value of a cluster, also called cluster center mu j (ii) a Wherein the value of K is a positive integer greater than 0;
step 1.4: for each feature map x, calculating x separately n With cluster center μ j Then obtaining the nearest cluster center, and determining the cluster mark of x, namely: c i =argmin j∈[1,m] ||x n -μ j | |, finally the sample x is classified into the corresponding cluster C i,x =C i U is x; wherein m represents the number of clusters;
step 1.5: updating the cluster center for each cluster C i Computing new cluster centersMu's' j ≠μ j Update mu j Is mu' j Otherwise mu j The change is not changed;
step 1.6: repeat step 1.4 and step 1.5 to μ j And no updating, and finally dividing the feature mapping vector into clusters: c = { C 1 ,C 2 ,...,C m }。
And 2, step: respectively inputting the clustered images into a DCGAN network, and respectively generating unlabeled images;
and 3, step 3: performing label distribution on the label-free image generated in the step 2 by adopting a clustering label smooth normalized loss function (CLS) to obtain a labeled generated image;
and 4, step 4: fusing the collected Pedestrian image library with the generated image obtained in the step 3, and expanding the collected Pedestrian image library to obtain a new Pedestrian image library Peerseries 02;
the specific implementation of the step 4 comprises the following substeps:
step 4.1: inputting the feature vectors of each cluster into a DCGAN network respectively, wherein the DCGAN network consists of a generative model (obtaining data distribution) and a distinguishing model D (predicting whether the input is real or generated in G); wherein G is a simple neural network, a feature vector is input, and then a graph is generated as an output; d is also a simple neural network, inputs an image and then generates a confidence;
and 4.2: using a loss function L GAN Training the DCGAN network to obtain a generated image;
L GAN =logD(x)+log(1-D(G(z)))
wherein D (x) is a confidence coefficient, the value is [0-1], and G (z) is a generated image;
step 4.3: all the generated images are blended together and fused with the collected Pedestrian image library pedistrian 01 to serve as a new Pedestrian image library pedistrian 02.
And 5: dividing the Peer 02 level into p blocks, and inputting each block into the CNN network for feature extraction to obtain the local features of the image;
the specific implementation of the step 5 comprises the following substeps:
step 5.1: inputting a generated new training set in a ResNet50 network, forming a 3D tensor T through forward convolution, and dividing the tensor T into p horizontal strips; wherein the value of p is a positive integer greater than 0;
step 5.2: sampling tensor T space into p horizontal strips through a pooling layer, and averaging all column vectors in the same strip into a single partial-level column vector g;
step 5.3: reducing the dimension of the vector g through a convolution layer with the kernel size of 1 multiplied by 1, and finally respectively inputting the column vector h after dimension reduction of each vector g into a classifier;
step 5.4: connecting p h to form a final descriptor of the input image, and obtaining local features of the image;
and step 5.5: during training, each classifier predicts the identity of an input image, and a multi-loss optimization strategy is adopted, so that the softmax classifiers of p classes are used for training, and the loss function uses the cross entropy loss function in the step 1.1;
and 6: performing combined training on the CNN network by adopting a clustering label smooth normalized loss function CLS and a cross entropy loss function in the training process;
the specific implementation of the step 6 comprises the following substeps:
step 6.1: constructing a clustering label smooth normalization loss function CLS;
in order to solve the problem of new data label distribution, the invention invents a cluster label smooth normalized loss function (CLS) to train the CNN model according to the similarity of cluster samples during training.
According to the invention, a cluster labeling smooth normalization loss function (CLS) is constructed according to the similarity of cluster samples, firstly, a label smooth regularization LSRO loss function of an abnormal value is given:
wherein p (K | x) represents the prediction probability that input x belongs to class K, where K is the number of sample classes; t is a distinguishing parameter, for a real training image, T =0; for the generated image, T =1;
wherein z is i ,z k Respectively generating non-normalized probabilities of the ith and the K images by using K clustering;
given by N i Cluster of individual classes C i The generated image x, i belongs to [1,m ]],q g (k | x) is the one-hot category labeling probability of the image x, and the expression is as follows:
for all k, 0-1 coding is used, first when k belongs to cluster C i Let K =1, otherwise K =0, where K ∈ {1,2, 3.., K }, and then respectively match the number with the general classAnd dividing by K to obtain the normalized clustering class label of the generated sample as follows:
considering the valid correct (ground-route) distribution so thatSince the samples are from cluster C i And C is i Each class has similar characteristics, so uniform distribution is adoptedIndicating that the generated sample x belongs to C i Probability of each class, in z k,x Denotes the log probability of unnormalization, let N i As a cluster C i Total number of classes of (2), then network normalized output z' k The expression of (a) is:
deriving partial prediction probability expressionsQ is to be g '(k | x) and p' (k | x) are substituted into the LSRO loss function to obtain CLS as:
in contrast to the LSRO penalty function, the CLS is written as:
the expression of the true sample label q' (k) of the CLS obtained from the existing label smooth normalized LSR is:
wherein delta k,t =q g (k/x);
It can be seen from the observation that epsilon =0.1 in LSR, when k ≠ t, the LSRO penalty function gives a given label to the generated sampleEpsilon is 1; in the cluster label smooth normalized loss function (CLS) of the present invention
Step 6.2: and (3) training the CNN model by using the combination of the clustering label smooth normalized loss function CLS and the cross entropy loss function.
Step 6.3: during testing, the collected Pedestrian image library Pedestrian01 is input into the CNN model trained in the step 6.2, and an initial result arrangement table is obtained.
And 7: during testing, re-ranking Re-ranking optimization is adopted, and a pedestrian Re-identification result is output.
The specific implementation of step 7 comprises the following substeps:
step 7.1: the invention reorders the pictures in the Pedestrian image Pedestrian01 to be detected by adopting the existing k-order derivative coding-based method, so that the identification result is improved. Our goal is to reorder the initial ordered list in step 6.3 so that more positive samples appear at the front of the list. First, k-th order neighbor (k-nn), i.e., the first k samples of the ordered list, is defined:
whereinMahalanobis distance of feature vectors 0 to k, p is the query image,k is the k-th order inverse coded feature vector and N () is the set. Then, k-order nearest neighbor (k-rnn) is defined
Wherein g is i For the ith image in Pedestrian01,is p and g i Are all in the k-th neighborhood of each other. However, due to a range of variations in illumination, pose, view angle, etc., positive samples may be excluded from the k-nn list, and we therefore employ a more robust k-rnn set.
I.e. for a collection of scriptsFind their k-rnn set for each sample q in (1)If the number of coincident samples reaches a certain condition, the coincident samples are mergedIn this way, the original text is notPositive samples in the setAnd bringing back again.
Step 7.2: calculating the Jaccard distance between the two images through the k-rnn set obtained in the step 7.1:
where p is the query image, g i For the ith image in Pedestrian01, k is the k-th order reciprocal coded feature vector.
Step 7.3: and (4) rearranging the initial result obtained in the step (6.3) according to the Jaccard distance obtained in the step (7.2), and finally outputting a pedestrian re-identification result.
It should be understood that parts of the specification not set forth in detail are well within the prior art.
It should be understood that the above description of the preferred embodiments is given for clarity and not for any purpose of limitation, and that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (6)
1. A pedestrian re-identification method based on clustering and block feature extraction is characterized by comprising the following steps:
step 1: collecting Pedestrian images under a monitoring camera to obtain a Pedestrian image library Pedestrian01, and carrying out k-means clustering on the Pedestrian 01;
step 2: respectively inputting the clustered images into a DCGAN network, and respectively generating unlabeled images;
and step 3: performing label distribution on the label-free image generated in the step 2 by adopting a clustering label smooth normalized loss function CLS to obtain a labeled generated image;
and 4, step 4: fusing the collected Pedestrian image library with the generated image obtained in the step 3, and expanding the collected Pedestrian image library to obtain a new Pedestrian image library Pedestrian02;
and 5: dividing the Peer 02 into p blocks horizontally, and inputting each block into a CNN network for feature extraction to obtain local features of the image;
step 6: performing combined training on the CNN network by adopting a clustering label smooth normalized loss function CLS and a cross entropy loss function in the training process;
and 7: during testing, re-ranking Re-ranking optimization is adopted, and a pedestrian Re-identification result is output.
2. The pedestrian re-identification method based on clustering and block feature extraction as claimed in claim 1, wherein the specific implementation of step 1 comprises the following sub-steps:
step 1.1: inputting Pedestrian image libraries Pedestrian01 to ResNet50 network under a monitoring camera, and adopting a cross entropy loss functionTraining to obtain a feature extraction model; wherein y is the original actual output,is the desired output;
step 1.2: inputting the training data set into the feature extraction model in step 1.1, and extracting the feature mapping vector x of the last convolutional layer n ;
Step 1.3: randomly selecting K feature mapping objects, wherein each object represents an initial mean value of a cluster, also called a cluster center mu j (ii) a Wherein the value of K is a positive integer greater than 0;
step 1.4: for each feature map x, calculating x separately n With cluster center μ j Then obtaining the nearest cluster center, and determining the cluster mark of x, namely: c i =argmin j∈[1,m] ||x n -μ j | |, finally the sample x is sorted into the corresponding cluster C i,x =C i U is x; wherein m represents the number of clusters;
step 1.5: in update clusterHeart, for each cluster C i Computing new cluster centersMu's' j ≠μ j Update mu j Is mu' j Otherwise mu j The change is not changed;
step 1.6: repeat step 1.4 and step 1.5 to μ j And no updating, and finally dividing the feature mapping vector into clusters: c = { C 1 ,C 2 ,...,C m }。
3. The pedestrian re-identification method based on clustering and block feature extraction as claimed in claim 1, wherein the specific implementation of step 4 comprises the following sub-steps:
step 4.1: respectively inputting the feature vectors of each cluster into a DCGAN network, wherein the DCGAN network consists of a generation model G and a distinguishing model D; wherein G is a simple neural network, a feature vector is input, and then a graph is generated as an output; d is also a simple neural network, inputs an image and then generates a confidence;
step 4.2: using a loss function L GAN Training the DCGAN network to obtain a generated image;
L GAN =logD(x)+log(1-D(G(z)))
wherein D (x) is a confidence coefficient, the value is [0-1], and G (z) is a generated image;
step 4.3: all the generated images are blended together and fused with the collected Pedestrian image library pedistrian 01 to serve as a new Pedestrian image library pedistrian 02.
4. The pedestrian re-identification method based on clustering and block feature extraction as claimed in claim 1, wherein the concrete implementation of step 5 comprises the following sub-steps:
step 5.1: inputting a generated new Pedestrian image library Pedestrian02 in a ResNet50 network, forming a 3D tensor T through forward convolution, and horizontally dividing the tensor T into p horizontal strips; wherein the value of p is a positive integer greater than 0;
and step 5.2: sampling a tensor T space into p horizontal strips through a pooling layer, and averaging all column vectors in the same strip into a single partial-level column vector g;
step 5.3: reducing the dimension of the vector g through a convolution layer with the kernel size of 1 multiplied by 1, and finally respectively inputting the column vector h after dimension reduction of each vector g into a classifier;
step 5.4: connecting p h to form a final descriptor of the input image, and obtaining local features of the image;
step 5.5: during training, each classifier predicts the identity of the input image, and is trained with the softmax classifiers of p classes using the multi-loss optimization strategy, and the loss function uses the cross-entropy loss function in step 1.1.
5. The pedestrian re-identification method based on clustering and block feature extraction as claimed in claim 1, wherein the specific implementation of step 6 comprises the following sub-steps:
step 6.1: constructing a clustering label smooth normalization loss function CLS;
first the label smoothing regularization LSRO loss function for a given outlier:
wherein p (K | x) represents the prediction probability that input x belongs to class K, where K is the number of sample classes; t is a distinguishing parameter, for a collected pedestrian image, T =0; for the generated image, T =1;
wherein z is i ,z k Respectively using the non-normalized probability of the ith and the K images generated by the K clustering;
given by N i Cluster of individual classes C i The generated image x, i belongs to [1,m ]],q g (k | x) is the one-hot category labeling probability of the image x, and the expression is as follows:
for all k, 0-1 coding is used, first when k belongs to cluster C i When K =1, otherwise K =0, where K belongs to {1,2,3,.., K }, and then the numbers are divided by the total class K, respectively, to obtain a normalized clustering class label of the generated sample as:
with uniform distributionIndicating that the generated sample x belongs to C i Probability of each class, using z k,x Denotes the log probability of unnormalization, let N i As a cluster C i Total number of classes of (2), then network normalized output z' k The expression of (a) is:
deriving partial prediction probability expressionsQ is to be g '(k | x) and p' (k | x) are substituted into the LSRO loss function to obtain CLS as:
in contrast to the LSRO penalty function, the CLS is written as:
the expression of the real sample label q' (k) of the CLS obtained from the existing label smooth normalized LSR is:
wherein delta k,t =q g (k/x), ε =0.1 in LSR, when k ≠ t, the LSRO loss function gives a label to the generated sampleEpsilon is 1;
step 6.2: training the CNN model by using a clustering label smooth normalization loss function CLS and a cross entropy loss function;
step 6.3: during testing, the collected Pedestrian image library Pedestrian01 is input into the trained CNN model in the step 6.2, and an initial result arrangement table is obtained.
6. The pedestrian re-identification method based on clustering and block feature extraction according to any one of claims 1 to 5, wherein the specific implementation of step 7 comprises the following sub-steps:
step 7.1: reordering the pictures in the acquired Pedestrian image Pedestrian01 to be detected by adopting a k-order derivative coding-based mode, so that the identification result is improved;
first, k-th order neighbor k-nn is defined, i.e. the first k samples of the ordered list:
whereinThe Mahalanobis distance from the feature vector 0 to k is obtained, p is the query image, k is the k-order reciprocal coded feature vector, and N () is a set;
next, k-order nearest neighbors k-rnn are defined:
wherein g is i For the ith image in Pedestrian01,is p and g i A set of conditions that are all neighbors of k-th order of the counterpart; however, due to a series of changes in illumination, attitude, and view angle, positive samples may be excluded from the k-nn list, and therefore a more robust k-rnn set is used:
i.e. for a collection of scriptsFind their k-rnn set for each sample q in the setIf the number of coincident samples reaches a certain condition, the coincident samples are combinedIn this way, the original place is notBringing back the positive samples in the set;
and 7.2: calculating the Jaccard distance between the two images through the k-rnn set obtained in the step 7.1:
where p is the query image, g i The ith image in Pedestrian01 is shown, and k is a feature vector coded by an inverse k order;
step 7.3: and (4) rearranging the initial result obtained in the step (6.3) according to the Jaccard distance obtained in the step (7.2), and finally outputting a pedestrian re-identification result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910243050.5A CN109961051B (en) | 2019-03-28 | 2019-03-28 | Pedestrian re-identification method based on clustering and block feature extraction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910243050.5A CN109961051B (en) | 2019-03-28 | 2019-03-28 | Pedestrian re-identification method based on clustering and block feature extraction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109961051A CN109961051A (en) | 2019-07-02 |
CN109961051B true CN109961051B (en) | 2022-11-15 |
Family
ID=67025138
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910243050.5A Active CN109961051B (en) | 2019-03-28 | 2019-03-28 | Pedestrian re-identification method based on clustering and block feature extraction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109961051B (en) |
Families Citing this family (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110619264B (en) * | 2019-07-30 | 2023-06-16 | 长江大学 | Method and device for identifying microseism effective signals based on UNet++, and method and device for identifying microseism effective signals based on UNet++ |
CN110555390B (en) * | 2019-08-09 | 2022-09-09 | 厦门市美亚柏科信息股份有限公司 | Pedestrian re-identification method, device and medium based on semi-supervised training mode |
CN110633751A (en) * | 2019-09-17 | 2019-12-31 | 上海眼控科技股份有限公司 | Training method of car logo classification model, car logo identification method, device and equipment |
CN110688966B (en) * | 2019-09-30 | 2024-01-09 | 华东师范大学 | Semantic guidance pedestrian re-recognition method |
CN110796026A (en) * | 2019-10-10 | 2020-02-14 | 湖北工业大学 | Pedestrian re-identification method based on global feature stitching |
CN110728238A (en) * | 2019-10-12 | 2020-01-24 | 安徽工程大学 | Personnel re-detection method of fusion type neural network |
CN110968735B (en) * | 2019-11-25 | 2023-06-20 | 中国矿业大学 | Unsupervised pedestrian re-identification method based on spherical similarity hierarchical clustering |
CN113032553A (en) * | 2019-12-09 | 2021-06-25 | 富士通株式会社 | Information processing apparatus, information processing method, and computer program |
CN111274992A (en) * | 2020-02-12 | 2020-06-12 | 北方工业大学 | Cross-camera pedestrian re-identification method and system |
CN111461002B (en) * | 2020-03-31 | 2023-05-26 | 华南理工大学 | Sample processing method for thermal imaging pedestrian detection |
CN111666843B (en) * | 2020-05-25 | 2023-04-28 | 湖北工业大学 | Pedestrian re-recognition method based on global feature and local feature splicing |
CN111612100B (en) * | 2020-06-04 | 2023-11-03 | 商汤集团有限公司 | Object re-identification method, device, storage medium and computer equipment |
CN112070010B (en) * | 2020-09-08 | 2024-03-22 | 长沙理工大学 | Pedestrian re-recognition method for enhancing local feature learning by combining multiple-loss dynamic training strategies |
CN112784674B (en) * | 2020-11-13 | 2022-07-15 | 北京航空航天大学 | Cross-domain identification method of key personnel search system based on class center self-adaption |
CN112488035B (en) * | 2020-12-14 | 2024-04-26 | 南京信息工程大学 | Cross-domain pedestrian re-identification method based on antagonistic neural network |
CN112597871B (en) * | 2020-12-18 | 2023-07-18 | 中山大学 | Unsupervised vehicle re-identification method, system and storage medium based on two-stage clustering |
CN113096080B (en) * | 2021-03-30 | 2024-01-16 | 四川大学华西第二医院 | Image analysis method and system |
CN113378620B (en) * | 2021-03-31 | 2023-04-07 | 中交第二公路勘察设计研究院有限公司 | Cross-camera pedestrian re-identification method in surveillance video noise environment |
CN113239782B (en) * | 2021-05-11 | 2023-04-28 | 广西科学院 | Pedestrian re-recognition system and method integrating multi-scale GAN and tag learning |
CN113420639A (en) * | 2021-06-21 | 2021-09-21 | 南京航空航天大学 | Method and device for establishing near-ground infrared target data set based on generation countermeasure network |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103258212A (en) * | 2013-04-03 | 2013-08-21 | 中国科学院东北地理与农业生态研究所 | Semi-supervised integrated remote-sensing image classification method based on attractor propagation clustering |
CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
EP3399465A1 (en) * | 2017-05-05 | 2018-11-07 | Dassault Systèmes | Forming a dataset for fully-supervised learning |
-
2019
- 2019-03-28 CN CN201910243050.5A patent/CN109961051B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103258212A (en) * | 2013-04-03 | 2013-08-21 | 中国科学院东北地理与农业生态研究所 | Semi-supervised integrated remote-sensing image classification method based on attractor propagation clustering |
EP3399465A1 (en) * | 2017-05-05 | 2018-11-07 | Dassault Systèmes | Forming a dataset for fully-supervised learning |
CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
Also Published As
Publication number | Publication date |
---|---|
CN109961051A (en) | 2019-07-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109961051B (en) | Pedestrian re-identification method based on clustering and block feature extraction | |
Wu et al. | Deep learning-based methods for person re-identification: A comprehensive review | |
CN111126360B (en) | Cross-domain pedestrian re-identification method based on unsupervised combined multi-loss model | |
Ming et al. | Deep learning-based person re-identification methods: A survey and outlook of recent works | |
Wu et al. | A comprehensive study on cross-view gait based human identification with deep cnns | |
CN111259786B (en) | Pedestrian re-identification method based on synchronous enhancement of appearance and motion information of video | |
Jiang et al. | Recognizing human actions by learning and matching shape-motion prototype trees | |
CN110717411A (en) | Pedestrian re-identification method based on deep layer feature fusion | |
Wang et al. | A survey of vehicle re-identification based on deep learning | |
Yang et al. | Discovering motion primitives for unsupervised grouping and one-shot learning of human actions, gestures, and expressions | |
CN111666851B (en) | Cross domain self-adaptive pedestrian re-identification method based on multi-granularity label | |
CN111639564B (en) | Video pedestrian re-identification method based on multi-attention heterogeneous network | |
Tang et al. | Multi-modal metric learning for vehicle re-identification in traffic surveillance environment | |
CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
CN109447123B (en) | Pedestrian re-identification method based on label consistency constraint and stretching regularization dictionary learning | |
CN110728216A (en) | Unsupervised pedestrian re-identification method based on pedestrian attribute adaptive learning | |
CN111046732A (en) | Pedestrian re-identification method based on multi-granularity semantic analysis and storage medium | |
Yi et al. | Motion keypoint trajectory and covariance descriptor for human action recognition | |
CN112819065A (en) | Unsupervised pedestrian sample mining method and unsupervised pedestrian sample mining system based on multi-clustering information | |
CN111723773A (en) | Remnant detection method, device, electronic equipment and readable storage medium | |
CN112464730A (en) | Pedestrian re-identification method based on domain-independent foreground feature learning | |
Yi et al. | Mining human movement evolution for complex action recognition | |
Singh et al. | A comprehensive survey on person re-identification approaches: various aspects | |
CN105825201A (en) | Moving object tracking method in video monitoring | |
Behera et al. | Person re-identification: A taxonomic survey and the path ahead |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |