Pedestrian re-identification system and method based on Hash coding learning and application

Abstract

The invention discloses a pedestrian re-recognition system and method based on Hash coding learning and application. The system comprises a pedestrian feature extraction module, an uncertainty estimation and loss calculation module, a model training module and a retrieval module. The pedestrian feature extraction module comprises a Hash main network and a momentum auxiliary network which are consistent in network structure, and is used for performing feature extraction on an input pedestrian image and outputting corresponding activation features; the uncertainty estimation and loss calculation module is used for calculating bit-level uncertainty and image-level uncertainty based on the activation features and constructing a total loss function; the model training module is used for training and optimizing the Hash main network and the momentum auxiliary network based on the private training data set; and the retrieval module is used for carrying out offline library building on the pedestrian images in the base library and carrying out online retrieval on the to-be-queried images. According to the method, the storage pressure and the comparison speed pressure of the pedestrian re-recognition features can be remarkably reduced, and the storage and calculation efficiency of a pedestrian re-recognition application system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a pedestrian re-identification system and method based on hash code learning and application. BACKGROUND

[0002] Pedestrian re-identification is a popular research problem in the field of computer vision, and has a wide application prospect in intelligent monitoring, public safety and other fields. With the gradual expansion of the application scale of pedestrian re-identification, the number of pedestrian re-identification comparisons has shown an explosive growth, which has brought significant storage problems and comparison speed problems.

[0003] Hashing technology (Hashing) has attracted widespread attention in pedestrian re-identification applications due to its low storage cost and fast computing speed. Traditional hash methods rely on hand-designed features, while in recent years with the development of deep learning, deep hash learning technology has made significant progress. Deep hash technology aims to map high-dimensional image data to low-dimensional binary Hamming space, but due to the optimization method of neural network back propagation, there is a discrete optimization problem of binary hash code in deep hash learning.

[0004] In the prior art, recent deep hash learning schemes often solve this challenge through binary approximation, but this approximate optimization process will cause uncertainty between the true value and the binary value, thereby reducing the retrieval performance of pedestrian re-identification, and cannot meet the high requirements of storage efficiency, comparison speed and retrieval performance in actual applications. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a pedestrian re-identification system and method based on hash code learning and application, to solve the problems of large storage pressure, slow comparison speed and poor retrieval performance in pedestrian re-identification applications, and to improve the overall performance of the pedestrian re-identification application system.

[0006] In order to achieve the above purpose, the solution of the present application is: A pedestrian re-identification system based on hash code learning, comprising a pedestrian feature extraction module, an uncertainty estimation and loss calculation module, a model training module and a retrieval module; The pedestrian feature extraction module comprises a hash main network and a momentum auxiliary network, and the hash main network and the momentum auxiliary network have the same structure and are used to extract features from input pedestrian images and output corresponding activation features; The uncertainty estimation and loss calculation module is used to calculate the bit-level uncertainty and image-level uncertainty based on the activation features, and to construct a total loss function; The model training module is configured to train and optimize the hash main network and the momentum auxiliary network based on a private training data set. The retrieval module is configured to perform offline library construction on pedestrian images in a library and online retrieval on to-be-queried images.

[0007] Further, the hash main network and the momentum auxiliary network both adopt a ResNet-50 network structure; the hash main network is configured to perform feature extraction on an input pedestrian image, output a first feature vector, and obtain a first activated feature after being processed by a tanh activation function; the momentum auxiliary network is configured to perform feature extraction on the same pedestrian image, output a second feature vector, and obtain a second activated feature after being processed by the tanh activation function.

[0008] Further, the uncertainty estimation and loss calculation module calculates a bit-level uncertainty of a formula: ; wherein, , , is a given pedestrian image, ; is a first activated feature, is a second activated feature; is a tanh activation function, whose output , is a hash main network, is a momentum auxiliary network; A formula for calculating an image-level uncertainty is: ; wherein, d is a sample feature vector dimension, is the i-th component of a bit-level uncertainty. k

[0009] Further, the total loss function includes a triple embedding loss function, a bit-level quantization loss function of hash coding, and a consistency loss function; the triple embedding loss function adopts an adaptive weight based on uncertainty; the bit-level quantization loss function of hash coding is individually weighted for each hash bit; and the consistency loss function is configured to maintain the consistency of the hash main network and the momentum auxiliary network.

[0010] Further, the model training module adopts a stochastic gradient descent algorithm to optimize the parameters of the hash main network, and updates the parameters of the momentum auxiliary network through a momentum update formula, the momentum update formula being: ; wherein, is a parameter of the momentum auxiliary network, ​For the parameters of the hash main network, for The constants between.

[0011] A person re-identification method based on hash encoding learning includes the following steps: S1. Data preparation: Obtain the training dataset, the base dataset, and the images to be queried; adjust the size of the pedestrian images. S2. Construct a dual-network architecture for pedestrian feature extraction, and build a hash main network and a momentum auxiliary network with completely identical structures. S3. Uncertainty estimation and loss function construction: Calculate bit-level uncertainty and image-level uncertainty, and construct the total loss function; S4: Model training, using a PK sampling strategy to extract training batches, and training and optimizing the hash main network and momentum auxiliary network; S5: Offline database construction, performing feature extraction, binarization processing, and storage on pedestrian images in the base database dataset; S6: Online retrieval, extracts features from the query image, binarizes it, calculates and sorts the Hamming distances between the image and pedestrian images in the database, and outputs the retrieval results.

[0012] Furthermore, the specific method of step S2 is as follows: Input pedestrian image Category label is , obtain features , The dimension of the sample feature vector. For hashing the main network; tanh The activation function is Its output Then hash main network Output Similarly, momentum-assisted networks Its network structure is similar to the hash main network. Completely identical, momentum-assisted network The output is .

[0013] Furthermore, the specific steps of step S3 are as follows: S31. Calculating bit-level uncertainty for: ;in, ， , Given a pedestrian image, ; S32, Calculate image-level uncertainty for: ;in, The dimension of the sample feature vector; S33. For pedestrian re-identification tasks, triplet embedding loss function for: ; in, 、 、 These represent the corresponding pedestrian images. eigenvectors, 、 、 This indicates the category label corresponding to the sample. and Same type and with Different categories, It is a constant. Triple loss function The weights are determined using adaptive weights based on uncertainty; therefore, Represented as: ,in, Representative sample Image-level uncertainty, when its value is high, the triplet loss weights This will increase exponentially, forcing the network to focus on samples with higher uncertainty when optimizing backpropagation; S34. Bit-level quantization loss of hash encoding for: ;in, The dimension of the sample feature vector. It is a constant; , It is a symbolic function; The hash loss function is weighted individually for each hash bit, meaning that if the uncertainty of that bit is considered... If the value is large, the optimization effort is stronger, forcing the bit to complete binarization as soon as possible and reducing the loss of retrieval accuracy caused by quantization; To maintain the hash master network With momentum auxiliary network Consistency, define the consistency loss function. for: , It is a constant; Uncertainty loss function based on deep hashing coding for: , Hash Main Network All learnable parameters are Momentum-assisted network All learnable parameters are During the backpropagation optimization process, Optimization is performed using backpropagation based on stochastic gradient descent. It will not participate in gradient descent updates, but will be updated through the momentum update formula: ,in, For one The constants between.

[0014] Furthermore, the specific method for step S6 is as follows: S61. Perform offline feature extraction and database construction for all images in the base database, and construct the database for each pedestrian image in the base database. Input into the pre-trained hash master network In the process, obtain the output of the hash main network. ; S62, targeting For each element, use a symbolic function. Binarization is performed, that is ,in, ;Will In compact binary form, soon The bit hash code is compressed and stored as follows: Each byte is stored in a dedicated hash database and linked to the corresponding pedestrian image tag. Alternatively, the storage location can be associated; S63. When performing an online query, a query image is provided. Similarly, a query hash code is generated: Use Hamming distance to calculate the query image. With each image in the database The similarity, i.e.: ； in, This is an indicator function, where a value of 1 is returned if the condition within the parentheses is true, and 0 is returned if the condition is false. In the system implementation, to achieve fast target retrieval, bitwise operations are used instead of floating-point operations to calculate the Hamming distance, and the hash code is... The value is determined by Mapped to Then calculate using the following formula.

[0015] in, This represents the XOR operation. For counting function (Population Count) This is typically achieved by calling the CPU instruction set; S64. Sort the results from smallest to largest based on the calculated Hamming distance and output the sorting results.

[0016] An application of a pedestrian re-identification system based on hash coding learning is disclosed, which is then applied to the fields of intelligent monitoring, public safety, or personnel management.

[0017] By adopting the above scheme, the pedestrian re-identification system, method and application based on hash coding learning of the present invention introduces deep hash learning into the pedestrian re-identification task. By performing hash binary coding learning on the output features of deep neural networks, the continuous features in Euclidean space are discretized into a low-dimensional binary Hamming space, which significantly reduces the storage pressure and comparison speed pressure of pedestrian re-identification features, improves the storage and computing efficiency of pedestrian re-identification application system, and can meet the needs of large-scale pedestrian re-identification application scenarios. Detailed Implementation

[0018] To further explain the technical solution of the present invention, the present invention will be described in detail below through specific embodiments.

[0019] As the scale of pedestrian re-identification applications gradually expands, the number of pedestrian re-identification comparisons has exploded, leading to significant storage and comparison speed issues. This invention proposes a pedestrian re-identification system, method, and application based on hash coding learning. First, deep hash learning is introduced into the pedestrian re-identification task. By discretizing continuous features in Euclidean space into a low-dimensional binary Hamming space, the storage pressure and comparison speed pressure of pedestrian re-identification features are significantly reduced, improving the computational efficiency of the pedestrian re-identification application system. Second, to solve the discretization optimization problem of binary hash codes, this invention models the bit-level uncertainty in hash coding and proposes a discretization optimization method, which can effectively improve the retrieval performance of hash codes after deep hash learning for pedestrian re-identification tasks.

[0020] The present invention proposes a pedestrian re-identification system based on hash coding learning, which includes a pedestrian feature extraction module, an uncertainty estimation and loss calculation module, a model training module, and a retrieval module.

[0021] The pedestrian feature extraction module includes a hash main network and a momentum auxiliary network, which have identical structures and are used to extract features from the input pedestrian image and output corresponding activation features. Both the hash main network and the momentum auxiliary network adopt the ResNet-50 network structure. The hash main network is used to extract features from the input pedestrian image and output a first feature vector, which is then processed by the tanh activation function to obtain the first activation feature. The momentum auxiliary network is used to extract features from the same pedestrian image and output a second feature vector, which is then processed by the tanh activation function to obtain the second activation feature.

[0022] The uncertainty estimation and loss calculation module is used to calculate bit-level uncertainty and image-level uncertainty based on the activation features, and to construct a total loss function; the uncertainty estimation and loss calculation module calculates bit-level uncertainty. The formula is: ;in, , , Given a pedestrian image, ; As the first activation feature, This is the second activation feature; The tanh activation function is used. Its output , For the hash main network, For kinetic-assisted networks; formula for calculating image-level uncertainty. for: ;in, d The dimension of the sample feature vector. For bit-level uncertainty, the first k The total loss function comprises three components: a triplet embedding loss function, a hash-encoded bit-level quantization loss function, and a consistency loss function. The triplet embedding loss function employs adaptive weights based on uncertainty. The hash-encoded bit-level quantization loss function is weighted individually for each hash bit. The consistency loss function is used to maintain the consistency between the hash main network and the momentum auxiliary network.

[0023] The model training module is used to train and optimize the hash main network and momentum auxiliary network based on a private training dataset. The model training module uses stochastic gradient descent to optimize the parameters of the hash main network and updates the parameters of the momentum auxiliary network using a momentum update formula, which is: ;in, For the parameters of the momentum-assisted network, For the parameters of the hash main network, for The constants between.

[0024] The retrieval module is used to build an offline database of pedestrian images in the base database and to perform online retrieval of images to be queried.

[0025] This invention also proposes a person re-identification method based on hash coding learning, comprising the following steps: S1. Data Preparation: Obtain the training dataset, the base dataset, and the images to be queried. Adjust the size of the pedestrian images. After adjusting the size of the pedestrian images, perform preprocessing operations on the pedestrian images, including one or more combinations of normalization, random horizontal flipping, and random cropping.

[0026] S2. Construct a dual-network architecture for pedestrian feature extraction, and build a hash main network and a momentum auxiliary network with completely identical structures. Define the hash main network Given a pedestrian image Category tags are Features can be obtained , Let be the dimension of the sample feature vector. (Definition) tanh The activation function is Its output Then define the hash main network. Output Similarly, define a momentum-assisted network. Its network structure is similar to the hash main network. Completely identical, define momentum-assisted network The output is .

[0027] S3. Uncertainty estimation and loss function construction: Calculate bit-level uncertainty and image-level uncertainty, and construct the total loss function; S31. Calculating bit-level uncertainty for: ;in, ， , Given a pedestrian image, ; S32, Calculate image-level uncertainty for: ;in, The dimension of the sample feature vector; S33. For pedestrian re-identification tasks, triplet embedding loss function for: ; in, 、 、 These represent the corresponding pedestrian images. eigenvectors, 、 、 This indicates the category label corresponding to the sample. and Same type and with Different categories, It is a constant. Triple loss function The weights are determined using adaptive weights based on uncertainty; therefore, Represented as: ,in, Representative sample Image-level uncertainty, when its value is high, the triplet loss weights This will increase exponentially, forcing the network to focus on samples with higher uncertainty when optimizing backpropagation; S34. Bit-level quantization loss of hash encoding for: ;in, The dimension of the sample feature vector. This is a constant, manually set during algorithm training, and has no range requirements; , It is a symbolic function; The hash loss function is weighted individually for each hash bit, meaning that if the uncertainty of that bit is considered... If the value is large, the optimization effort is stronger, forcing the bit to complete binarization as soon as possible and reducing the loss of retrieval accuracy caused by quantization; To maintain the hash master network With momentum auxiliary network Consistency, define the consistency loss function. for: , It is a constant, manually set during algorithm training, and has no range requirements; The uncertainty loss function based on deep hashing is: , Hash Main Network All learnable parameters are Momentum-assisted network All learnable parameters are During the backpropagation optimization process, Optimization is performed using backpropagation based on stochastic gradient descent. It will not participate in gradient descent updates, but will be updated through the momentum update formula: ,in, For one The constants between.

[0028] S4: Model training, using a PK sampling strategy to extract training batches, and training and optimizing the hash main network and momentum auxiliary network.

[0029] S5: Offline database construction, performing feature extraction, binarization, and storage on pedestrian images in the base database dataset.

[0030] S6: Online retrieval, performing feature extraction and binarization on the query image, calculating and sorting the Hamming distances with pedestrian images in the base database, and outputting the retrieval results. Specifically, it includes the following steps: S61. Perform offline feature extraction and database construction for all images in the base database, and construct the database for each pedestrian image in the base database. Input into the pre-trained hash master network In the process, obtain the output of the hash main network. ; S62, targeting For each element, use a symbolic function. Binarization is performed, that is ,in, ;Will In compact binary form, soon The bit hash code is compressed and stored as follows: Each byte is stored in a dedicated hash database and linked to the corresponding pedestrian image tag. Alternatively, the storage location can be associated; S63. When performing an online query, a query image is provided. Similarly, a query hash code is generated: Use Hamming distance to calculate the query image. With each image in the database The similarity, i.e.: ;in, This is an indicator function, where a value of 1 is returned if the condition within the parentheses is true, and 0 is returned if the condition is false. In the system implementation, to achieve fast target retrieval, bitwise operations are used instead of floating-point operations to calculate the Hamming distance, and the hash code is... The value is determined by Mapped to Then calculate using the following formula. ;in, This represents the XOR operation. For counting function (Population Count) This is typically achieved by calling the CPU instruction set; S64. Sort the results from smallest to largest based on the calculated Hamming distance and output the sorting results.

[0031] This invention introduces deep hash learning into the pedestrian re-identification task. By performing hash binary encoding learning on the output features of deep neural networks in the pedestrian re-identification task, the storage pressure and comparison speed pressure of pedestrian re-identification features can be significantly reduced, thereby improving the storage and computing efficiency of the pedestrian re-identification application system.

[0032] This invention models the bit-level uncertainty in hash coding and proposes a discrete optimization method that can effectively improve the retrieval performance of hash coding after deep hash learning for pedestrian re-identification tasks.

[0033] The following uses a dataset to verify the pedestrian re-identification method based on hash encoding learning of this invention.

[0034] This invention uses a private dataset as training data, which contains 28,358 people and 242,584 pedestrian digital images.

[0035] First, this invention uses the ResNet-50 network structure as the pedestrian image feature network for experiments, i.e., the hash main network. and momentum auxiliary network All networks used are ResNet-50. Using a private training dataset and the loss function proposed in step S3, stochastic gradient descent is employed to optimize the hash master network. and momentum auxiliary network Optimization is performed to ultimately obtain the optimal hash master network. The images were then used for testing. During the optimization process, pedestrian images captured by different cameras were uniformly resized to 256x128 pixels. A p-sampling strategy (i.e., randomly selecting P pedestrians, and extracting K images from each person) was adopted to form a training batch. Specifically, in the experiment of this invention, each training batch used 32 pedestrians, with 8 images used for each person for optimization. The initial learning rate for optimization was 0.01, and stochastic gradient descent was used for optimization, with the learning rate gradually decreasing to 0.0001. The constant in the triplet loss... The coefficients in the momentum update formula Parameters in the loss function , .

[0036] This invention uses two publicly available test sets, Market-1501 and DukeMTMC-reID, to compare with four other existing pedestrian re-identification models in terms of Rank-1 priority and Mean Average Precision (mAP).

[0037] Table 1: Comparison results on the test set Market-1501.

[0038]

[0039] Table 2: Comparison results on the test set DukeMTMC-reID.

[0040]

[0041] The test results show that the method of the present invention has good performance in pedestrian re-identification tasks and can meet the needs of practical applications.

[0042] This invention models the bit-level uncertainty in hash coding and proposes a discrete optimization method. By constructing a reasonable loss function, the network can focus on samples and bits with high uncertainty during training, effectively solving the discrete optimization problem of binary hash codes and improving the retrieval performance of hash codes after deep hash learning for person re-identification tasks. On the public test set Market-1501, the Rank-1 reaches 94.2% and the mAP reaches 87.2%; on the DukeMTMC-reID test set, the Rank-1 reaches 88.1% and the mAP reaches 78.6%, demonstrating good retrieval performance.

[0043] This invention employs a dual-network architecture of a hash master network and a momentum auxiliary network, maintaining consistency between the two networks through a consistency loss function. This provides a reliable foundation for uncertainty estimation and further enhances the model's stability and reliability. During the retrieval process, bitwise operations are used instead of floating-point operations to calculate the Hamming distance, significantly improving the speed of similarity comparison and enabling rapid output of retrieval results to meet the demands of real-time applications.

[0044] The system and method of this invention have a wide range of applications and can be applied to multiple fields such as intelligent monitoring, public safety, and personnel management, and have high practical value and promising prospects for promotion.

[0045] The above embodiments are not intended to limit the product form and style of the present invention. Any appropriate changes or modifications made by those skilled in the art should be considered as not departing from the patent scope of the present invention.

Claims

1. A pedestrian re-identification system based on hash coding learning, characterized in that, The application relates to a pedestrian feature extraction module, an uncertainty estimation and loss calculation module, a model training module and a retrieval module. The pedestrian feature extraction module comprises a hash main network and a momentum auxiliary network, the hash main network and the momentum auxiliary network are completely consistent in structure, are used for extracting features of input pedestrian images and output corresponding activation features, and are used for extracting features of the same pedestrian images. The uncertainty estimation and loss calculation module is used for calculating bit-level uncertainty and image-level uncertainty based on the activation features and constructing a total loss function. The model training module is used for training and optimizing the hash main network and the momentum auxiliary network based on a private training data set. The retrieval module is used for offline database construction of pedestrian images in a base database and online retrieval of to-be-queried images. The hash main network and the momentum auxiliary network both adopt a ResNet-50 network structure; the hash main network is used for extracting features of input pedestrian images, outputting a first feature vector and obtaining a first activation feature after tanh activation function processing; and the momentum auxiliary network is used for extracting features of the same pedestrian images, outputting a second feature vector and obtaining a second activation feature after tanh activation function processing. The uncertainty estimation and loss calculation module calculates the bit-level uncertainty The formula is: ; wherein, , , for a given pedestrian image, ; is a first activation function, is a second activation function; is a tanh activation function, whose output , is a hash main network, is a kinetic auxiliary network; Formula for calculating image-level uncertainty for: ;in, d The dimension of the sample feature vector. For bit-level uncertainty, the first k Each component.

2. The pedestrian re-identification system based on hash coding learning of claim 1, wherein: The total loss function comprises a ternary embedding loss function, a hash coding bit-level quantization loss function and a consistency loss function; the ternary embedding loss function adopts an adaptive weight based on uncertainty; the hash coding bit-level quantization loss function is individually weighted for each hash bit; and the consistency loss function is used for keeping the consistency of the hash main network and the momentum auxiliary network.

3. The system for pedestrian re-identification based on hash coding learning as claimed in claim 1, wherein: The model training module adopts a stochastic gradient descent algorithm to optimize parameters of the hash main network and updates parameters of the momentum auxiliary network through a momentum update formula, the momentum update formula being as follows:

4. The system for pedestrian re-identification based on hash coding learning as claimed in claim 1, wherein: The application further relates to a pedestrian retrieval method comprising the following steps: ; wherein, are parameters of the momentum-augmented network, are parameters of the hash main network, are a constant between and 5. A pedestrian re-identification method based on hash coding learning, characterized in that, S1, data preparation: obtaining a training data set, a base database data set and to-be-queried images, and adjusting the size of pedestrian images; S2, constructing a pedestrian feature extraction double-network architecture, and building a hash main network and a momentum auxiliary network which are completely consistent in structure; S3, uncertainty estimation and loss function construction, calculating bit-level uncertainty and image-level uncertainty, and constructing a total loss function, the specific steps being as follows: The hash loss function is individually weighted for each hash bit. S31, calculate the uncertainty at the bit level is: ; wherein, ， , for a given image of pedestrians, ; S32, calculating image-level uncertainty is: ; wherein, is a sample feature vector dimension; S33、for pedestrian re-identification task, triple embedding loss function is: ； in, 、 、 These represent the corresponding pedestrian images. eigenvectors, 、 、 This indicates the category label corresponding to the sample. and Same type and with Different categories, It is a constant. Triple loss function The weights are determined using adaptive weights based on uncertainty; therefore, Represented as: ,in, Representative sample Image-level uncertainty; S34, loss of bit-level quantization of hash encoding is: ; wherein, is a sample feature vector dimension, is a constant; , is a sign function; The uncertainty loss function based on deep hash coding is as follows: To maintain the consistency of the hash main network with the momentum auxiliary network , a consistency loss function is defined as: , is a constant; S4, model training, adopting a PK sampling strategy to extract a training batch, and training and optimizing the hash main network and the momentum auxiliary network; ， Hash main network all learnable parameters of the hash main network all learnable parameters of the momentum auxiliary network all learnable parameters of the momentum auxiliary network during the backpropagation optimization process, optimization using backpropagation based on stochastic gradient descent, then do not participate in the gradient descent update, but are updated by the momentum update formula: wherein is a constant between one and one. S5, offline database construction, extracting features of pedestrian images in the base database data set, performing binary processing and storing the pedestrian images; S6, online retrieval, extracting features of to-be-queried images, performing binary processing, calculating Hamming distances of the to-be-queried images from pedestrian images in the base database and sorting the Hamming distances, and outputting retrieval results. The specific method of step S2 is as follows:

6. The method of claim 5, wherein the hash code learning based pedestrian re-identification method is characterized by, The specific method of step S6 is as follows: Inputting pedestrian image , the category label is , obtaining features , is the dimension of the sample feature vector, is a hash main network; the tanh activation function is , the output of which is ; the output of the hash main network is ; similarly, the momentum auxiliary network , the network structure of which is completely consistent with the hash main network , the output of the momentum auxiliary network is .

7. The method of claim 5, wherein the hash coding learning based pedestrian re-identification method is characterized by, Population Count S61, offline feature extraction and library building operation are performed on all images in the library, and each pedestrian image in the library is input into the already trained hash master network , and the output of the hash master network is obtained ; S62, for each element of the sign function is performed, i.e. wherein, ; will be stored in a compact binary form, i.e. the bit hash code will be stored in a compact binary form, i.e. the bit hash code will be stored in a compact binary form, i.e. the bit hash code will be stored in a compact binary form, i.e. the bit hash code will be stored in a compact binary form, i.e. the S63, when querying online, a given image is queried , and a query hash code is also generated: ; using hamming distance to calculate the similarity of the image to be queried with each image in the base library the similarity of each image in the base library​ ； Wherein, is an indicator function, that is, the condition in the parentheses is 1, and the condition in the parentheses is 0. In system implementation, in order to realize fast retrieval of the target, bit operation is used instead of floating point operation to calculate the Hamming distance, and the value of the hash code is mapped from to ; then the formula is used to calculate wherein denotes an exclusive OR operation, is a counting function S64, sorting the calculated Hamming distances from small to large and outputting the sorting results. implemented by invoking the CPU instruction set; ​ 8. An application of a pedestrian re-identification system based on hash coding learning, characterized in that: The pedestrian re-identification system based on hash coding learning according to any one of claims 1-4 is applied to the field of intelligent monitoring, public security or personnel management.

Images