A pedestrian re-identification method

Abstract

The present invention discloses a pedestrian re-identification method, which belongs to the field of computer vision. The method adopts a recognition terminal and a recognition system for the recognition terminal; the method comprises the following steps: Step 1: Prepare a training data set: Prepare Market‑1501, Dukemtmc‑Reid and Cuhk03 data sets, and train the data sets. Before training, preprocess the data sets, including image preprocessing and data enhancement; Step 2: Establish a posture estimation algorithm and perform training and optimization. The posture estimation algorithm is used to model and match the appearance under different postures, realize domain adaptive pedestrian re-identification, achieve better recognition effect and robustness, and can make a portable pedestrian re-identification device. The device is used to perform stable and high-accuracy identity recognition of pedestrians in real-time video streams, which can meet people's needs for pedestrian re-identification in various scenarios, and has high practicality and application prospects.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a pedestrian re-identification method. Background Art

[0002] Security is vital to the survival and development of society and enterprises. Especially in today's highly developed modern technology, criminal methods have become more intelligent and covert, so it is more urgent to strengthen modern security technology.

[0003] People have an increasing demand for security monitoring and crowd counting in their daily lives. Traditional monitoring systems can only recognize the general outline of a person, and it is difficult to recognize the identity of pedestrians. However, pedestrian re-identification technology can accurately identify pedestrians in different scenarios and has broad application prospects.

[0004] After searching, the Chinese patent application number CN115410230A discloses a pedestrian re-identification method and device, which discloses a device, method, electronic device and computer-readable storage medium to solve the problems of poor generalization ability and low cross-domain accuracy of pedestrian re-identification models;

[0005] The Chinese patent application number CN114202739A discloses a method for video pedestrian re-identification, which discloses a solution for extracting vital signs by finding other videos of the same pedestrian that match a given video of a pedestrian to be identified in a video collection.

[0006] The above schemes all disclose solutions, and also mention the problems existing in existing pedestrian re-identification systems: Since pedestrians' appearance may change in different scenarios, such as wearing different clothes, hats or masks, these changes may affect the accuracy of pedestrian re-identification. Appearance may change under different viewing angles, and traditional pedestrian re-identification technology often cannot handle such changes. Postures and behaviors in different scenarios may also be different, which will bring certain challenges to pedestrian re-identification.

[0007] Especially for cross-domain problems, the performance of pedestrian re-identification systems in different scenarios varies greatly, so cross-domain training and testing are required, but cross-domain problems are still a challenge. How to improve recognition effects and robustness has high practical value in practical applications. In addition, with the development of science and technology today, there are still many small devices or corresponding components with sufficient computing power and functions. How to combine recognition devices to meet the needs of portable, efficient and accurate pedestrian re-identification is also an urgent problem to be solved. Summary of the invention

[0008] The purpose of the present invention is to solve the defects in the prior art and propose a pedestrian re-identification method.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A pedestrian re-identification method, the method using an identification terminal and an identification system for the identification terminal;

[0011] The method comprises the following steps:

[0012] Step 1: Prepare training dataset:

[0013] Prepare the Market-1501, Dukemtmc-Reid and Cuhk03 datasets and train them. Before training, preprocess these datasets, including image preprocessing and data enhancement.

[0014] Step 2: Establish a pose estimation algorithm and perform training and optimization:

[0015] During training, select and tune hyperparameters including learning rate, loss function, and regularization term, and use cross-validation for model selection and performance evaluation;

[0016] Step 3: Implement the pose estimation algorithm:

[0017] Use deep learning posture estimation algorithms to build posture estimation models and match the appearance of pedestrians in different postures to estimate the posture information of pedestrians;

[0018] Step 4: Adjust and optimize the pose estimation module:

[0019] Training, evaluating and optimizing the posture evaluation model, inputting the feature representation extracted by the CNN model and the current posture parameters into the posture estimation module, obtaining new posture parameters, and updating the optimized objective function, iterating the process until convergence, that is, the optimized objective function reaches a stable state;

[0020] Step 5: Get the recognition results and transfer the model as needed:

[0021] Transplant the algorithm and training model into the recognition terminal, and realize the recognition and communication transmission of the device.

[0022] Furthermore, the identification terminal is worn on any compatible part of the human body through a strap, and the identification terminal is any terminal device with computing and communication capabilities;

[0023] The identification terminal includes a terminal interconnection communication module, a data feedback module, a consensus communication module, a terminal feedback confirmation module, a local operation module, a terminal independent training module, a data interaction synchronization module, a posture evaluation module and a data classification module.

[0024] The terminal interconnection communication module, the data feedback module and the consensus communication module are all connected to the terminal feedback confirmation module and the local operation module, the terminal feedback confirmation module and the local operation module are all connected to the terminal independent training module and the data interaction synchronization module, and the terminal independent training module and the data interaction synchronization module are all connected to the posture assessment module and the data classification module.

[0025] Furthermore, the terminal interconnection communication module is used to realize data communication with the cloud and data communication between multiple terminals via wireless means;

[0026] The data feedback module is used to obtain the data and terminal information and obtain the corresponding data feedback through the cloud;

[0027] The consensus communication module is used to issue consensus logic instructions for unified execution to the terminal devices within the range according to the data, type, location, environment and current status information of the connected terminal devices;

[0028] The terminal feedback confirmation module is used to obtain the terminal location information, wearing location information, information disclosed by the wearing user, environmental weather information and terminal density information through the feedback information obtained from the terminal feedback and network communication;

[0029] The local operation module is used to call the local computing component to perform local operation, storage and training of data according to the consensus logic instructions;

[0030] The terminal independent training module is used to obtain the training model, training method and training data adapted to the data received by the current terminal through the local operation module and data communication, and perform independent operation training on such data through the computing component;

[0031] The data interaction and synchronization module is used to synchronize and interact the local information and operation information of the terminal according to logic. At the same time, it is also used to communicate data among multiple compatible terminals within the range and share data and computing power;

[0032] A data classification module is used to classify the acquired data, and store and operate the classified data as a data set;

[0033] The posture evaluation module is used to calculate and analyze the posture of the current terminal wearer and communicate the evaluation logic and evaluation results.

[0034] Furthermore, step 1 specifically includes the following steps:

[0035] Step 1.1: Download Market-1501, Dukemtmc-Reid, and Cuhk03 datasets.

[0036] Step 1.2: Image preprocessing:

[0037] For each image, cropping and scaling are required to unify the image size to 256*128 pixels. At the same time, image normalization is required to map pixel values ​​to the range of [0,1].

[0038] Step 1.3: Data Augmentation:

[0039] To increase the diversity of training data, each image requires random horizontal flipping, random cropping, and other data augmentation operations.

[0040] Furthermore, step 2 specifically includes the following steps:

[0041] Step 2.1: Feature Extraction:

[0042] Use Resnet-50 as the backbone network to extract the feature vector of pedestrian images;

[0043] Step 2.2: Feature vector fusion:

[0044] The feature vectors are fused using the feature distance fusion method to obtain a more comprehensive and accurate representation of pedestrian features;

[0045] Step 2.3: Person Re-identification:

[0046] The pedestrian feature vector is classified using the SoftMax classifier to obtain the identity of the pedestrian.

[0047] Furthermore, step 3 specifically includes the following steps:

[0048] Step 3.1: Feature Extraction:

[0049] Input pedestrian images into the pre-trained CNN model to extract feature representations of the images;

[0050] Step 3.2: Posture information evaluation:

[0051] Evaluate pedestrian posture information;

[0052] Step 3.3: Iterative optimization:

[0053] The pose estimation problem is converted into an optimization problem and the pose parameters of the pedestrian are optimized iteratively.

[0054] Furthermore, step 4 specifically includes the following steps:

[0055] Step 4.1: Pose Estimation Model Training:

[0056] The MPII Human Pose Dataset, an annotated human pose dataset, is used for model training.

[0057] Step 4.2: Pose Estimation Model Evaluation:

[0058] After training, the model is evaluated using a test dataset, the predictions are compared with the annotated data, and the error is calculated;

[0059] Step 4.3: Application of pose estimation model:

[0060] Apply the trained pose estimation model to person re-identification and construct pedestrian feature representation by predicting pedestrian pose information;

[0061] Step 4.4: Integration of feature representations:

[0062] The feature representations obtained from person re-identification and pose estimation models are combined into a single feature representation. This is achieved by concatenating the feature vectors obtained from the two models into a single vector. The resulting feature representation is used for matching and retrieval tasks.

[0063] Step 4.5: Evaluation:

[0064] The performance of the proposed method is evaluated on benchmark datasets using standard evaluation metrics.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The posture estimation algorithm is used to model and match the appearance under different postures, realizing domain adaptive pedestrian re-identification and achieving better recognition effect and robustness;

[0067] It can achieve significant performance improvement when dealing with visual differences between different domains, and realize more accurate and robust pedestrian re-identification. In addition, the algorithm and training model can be transplanted to the corresponding terminal device, and a portable pedestrian re-identification device can be made. The device can be used to perform stable and highly accurate identity recognition of pedestrians in real-time video streams. It also has the characteristics of portability and low power consumption, which improves the accuracy and stability of recognition, and can achieve comprehensiveness and efficiency of data through data communication, and can meet people's needs for pedestrian re-identification in various scenarios. It has high practicality and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0069] Figure 1 This is a flow chart of the pedestrian re-identification method proposed by the present invention;

[0070] Figure 2 A schematic diagram of the flow of the recognition system of the pedestrian re-identification method proposed in the present invention;

[0071] Figure 3 A schematic diagram of the implementation logic in the embodiment of the present invention;

[0072] Figure 4 This is a schematic diagram of terminal wearing identification in an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0074] Embodiment 1

[0075] Reference Figure 1 , a pedestrian re-identification method, the method using an identification terminal and an identification system for the identification terminal;

[0076] The method comprises the following steps:

[0077] Step 1: Prepare training dataset:

[0078] Prepare the Market-1501, Dukemtmc-Reid and Cuhk03 datasets and train them. Before training, preprocess these datasets, including image preprocessing and data enhancement.

[0079] It should be noted that image preprocessing: for each image, it is necessary to crop and scale the image to a uniform size of 256*128 pixels. At the same time, the image needs to be normalized to map the pixel values ​​to the range of [0,1].

[0080] Data enhancement: In order to increase the diversity of training data, it is necessary to perform data enhancement operations such as random horizontal flipping and random cropping on each image.

[0081] Step 2: Establish a pose estimation algorithm and perform training and optimization:

[0082] During training, select and tune hyperparameters including learning rate, loss function, and regularization term, and use cross-validation for model selection and performance evaluation;

[0083] Furthermore, the DG-Net network structure implementation process includes the following components:

[0084] 1. Backbone Network:

[0085] The backbone network is responsible for extracting the feature representation of pedestrian images. The backbone network is trained on the pre-trained image classification task and then applied to the pedestrian re-identification task through transfer learning.

[0086] ResNet-50 was selected as the basic network architecture, and the top-level classifier was removed. In order to retain more image features, the convolutional layer and pooling layer were retained, and the image features were converted into a fixed-length vector representation. The parameters of the pre-trained basic network were initialized on the Market1501 dataset, and end-to-end training was performed to optimize the loss function, adjust the network parameters, and perform fine-tuning as needed.

[0087] 2. Person Identification Branch:

[0088] An additional global average pooling layer is added for feature extraction. The output feature map of the backbone network has a size of H×W×C, where H and W are the height and width of the feature map, and C is the number of channels. The output feature map of the backbone network is input into the global average pooling layer. The function of the global average pooling layer is to reduce the spatial dimension of the feature map to 1, retain the channel dimension, and average the feature map of each channel to obtain a C-dimensional global feature vector.

[0089] Finally, the global feature vector of the pedestrian image is output. The length of the global feature vector is equal to the number of channels C, which can be regarded as a point in the pedestrian image in the feature space, which has information representing the identity of the pedestrian.

[0090] Through the above operations, the pedestrian recognition branch can extract the global features of the pedestrian image and obtain a global feature vector that represents the pedestrian's identity information. This global feature vector can be used for feature matching and similarity calculation in the pedestrian re-identification task.

[0091] 3. Local feature extraction:

[0092] In order to capture the local features of pedestrian images, a local feature extraction module is introduced. This module divides the pedestrian image into multiple local regions, such as the upper body and the lower body. Then, feature extraction is performed on each local region to obtain a local feature vector.

[0093] The pedestrian image is divided into upper body and lower body by horizontal segmentation. The features of each local area are extracted to obtain the local feature vector. The feature extraction is performed using a convolutional neural network architecture similar to the backbone network. All local feature vectors are combined to form a comprehensive feature representation of the pedestrian. The local feature vectors are combined by weighted averaging. The combined feature vector retains the local feature information of each local area.

[0094] Through the above operations, the local feature extraction module extracts the feature vector of each local area from the pedestrian image to capture the local feature information of the pedestrian. These local feature vectors are fused with the global feature vector.

[0095] 4. Graph Neural Network:

[0096] It is used to fuse and associate global features with local features. It adopts the idea of ​​graph model and represents global features and local features as a graph structure, where each node represents a feature vector. By performing information transmission and aggregation operations in the graph, the mutual influence and fusion of global features and local features are achieved.

[0097] To further illustrate, building a graph structure is to represent the global features and local features as a graph structure, where each node represents a feature vector, the global features and local features can be used as node features respectively, and the topological structure of the graph is represented by an adjacency matrix or an edge list;

[0098] In addition, the information transmission and aggregation of graph neural networks perform information transmission and aggregation operations in the graph to achieve the mutual influence and fusion of global features and local features.

[0099] Graph neural networks achieve information transfer by aggregating information from neighboring nodes of a node and updating the feature representation of the node. Multiple layers of graph neural networks are used to perform multiple information transfers and aggregations to enhance the expressiveness and robustness of features.

[0100] Feature fusion and output,In the last layer of the graph neural network, the weighted average method is used to,fuse the global features and local features to obtain a comprehensive feature,representation.

[0101] The final feature representation is used for feature matching and similarity calculation in the pedestrian re-identification task.

[0102] Step 3: Implement the pose estimation algorithm:

[0103] Use deep learning posture estimation algorithms to build posture estimation models and match the appearance of pedestrians in different postures to estimate the posture information of pedestrians;

[0104] Step 4: Adjust and optimize the pose estimation module:

[0105] Training, evaluating and optimizing the posture evaluation model, inputting the feature representation extracted by the CNN model and the current posture parameters into the posture estimation module, obtaining new posture parameters, and updating the optimized objective function, iterating the process until convergence, that is, the optimized objective function reaches a stable state;

[0106] Step 5: Get the recognition results and transfer the model as needed:

[0107] Transplant the algorithm and training model into the recognition terminal, and realize the recognition and communication transmission of the device.

[0108] In this embodiment, for the implementation of the posture estimation algorithm:

[0109] A deep learning-based pose estimation algorithm is used to model and match the appearance of pedestrians in different poses to improve the accuracy and robustness of pedestrian re-identification.

[0110] The specific implementation method is to use a deep neural network to train pedestrian images and learn the appearance features and posture information of pedestrians. A method based on a convolutional neural network (CNN) is used to extract features and estimate postures of pedestrian images, specifically:

[0111] First, we input pedestrian images into a pre-trained CNN model to extract feature representations of the images.

[0112] Then, a posture estimation module is used to estimate the posture information of the pedestrian, where:

[0113] Input feature representation: The input of this module is the feature representation extracted from the pedestrian image. These features can be extracted from the pedestrian image by the CNN model and processed by the convolution layer and the pooling layer.

[0114] Pose parameter estimation: Based on the input feature representation, the pose parameters of the pedestrian are estimated using a pose estimation algorithm. The pose parameters include three-dimensional rotation angles and translation vectors. The pose estimation algorithm is based on a convolutional neural network (CNN).

[0115] Attitude parameter optimization: After obtaining the attitude parameters, nonlinear optimization methods are used to optimize them to obtain more accurate results.

[0116] Output posture information: The final output posture information includes the pedestrian's three-dimensional rotation angle and translation vector, which can be used for subsequent applications such as pedestrian posture recognition and pedestrian behavior analysis.

[0117] In order to improve the accuracy and robustness of pose estimation, an iterative optimization strategy is adopted.

[0118] It should be noted that the posture estimation problem is transformed into an optimization problem, and the posture parameters of the pedestrian are gradually optimized through iteration.

[0119] Use a gradient descent-based optimization algorithm to update and adjust the posture parameters;

[0120] In each iteration, the feature representation extracted by the CNN model and the current posture parameters are input into the posture estimation module to obtain new posture parameters and update the optimized objective function. The iteration process continues until convergence, that is, the optimized objective function reaches a stable state.

[0121] Through the above design, the appearance model of pedestrians in different postures can be obtained. By matching the model with the existing pedestrian features, the accuracy and robustness of pedestrian re-identification can be improved. This method has achieved good results in pedestrian re-identification and can adapt to pedestrian re-identification problems under different postures and lighting conditions.

[0122] It should be noted that the implementation of the posture estimation algorithm can be modified and optimized according to actual conditions. For example, a deeper CNN model can be used to extract more complex feature representations, or a more advanced optimization algorithm can be used to improve the accuracy and speed of posture estimation.

[0123] In this embodiment, the posture estimation module needs to be adjusted and optimized in detail to adapt to different data sets and application scenarios. The specific steps are as follows:

[0124] (1) Pose estimation model training

[0125] The MPII Human Pose dataset, a labeled human pose dataset, is used for model training. During the training process, the input image and pose annotation are used as the input of the network, and the predicted pose information is output. The network weights are optimized by back-propagating the error between the predicted result and the annotated data. The optimization goal is to minimize the Euclidean distance or angle error between the predicted result and the annotation.

[0126] (2) Pose Estimation Model Evaluation

[0127] After training is completed, the test data set is used to evaluate the model, compare the error between the prediction results and the labeled data, and evaluate the model performance by calculating indicators such as prediction accuracy, recall rate, and F1-score.

[0128] (3) Application of posture estimation model

[0129] The trained posture estimation model is applied to pedestrian re-identification. By predicting the pedestrian's posture information, a pedestrian posture space model is constructed, and then the appearance of pedestrians in different postures is modeled and matched, thereby improving the accuracy and robustness of pedestrian re-identification.

[0130] (4) Model evaluation and optimization

[0131] After training is completed, the model is evaluated and optimized. The performance of the model is evaluated using indicators such as accuracy, precision, recall, and F1 value. At the same time, in order to improve the performance of the model, data enhancement, model structure optimization, and hyperparameter adjustment are performed.

[0132] Embodiment 2

[0133] like Figure 2 and Figure 4 As shown, based on the first embodiment, the identification terminal is worn on any compatible part of the human body through a strap, and the identification terminal is any terminal device with computing and communication capabilities;

[0134] The identification terminal includes a terminal interconnection communication module, a data feedback module, a consensus communication module, a terminal feedback confirmation module, a local operation module, a terminal independent training module, a data interaction synchronization module, a posture evaluation module and a data classification module.

[0135] The terminal interconnection communication module, the data feedback module and the consensus communication module are all connected to the terminal feedback confirmation module and the local operation module, the terminal feedback confirmation module and the local operation module are all connected to the terminal independent training module and the data interaction synchronization module, and the terminal independent training module and the data interaction synchronization module are all connected to the posture assessment module and the data classification module.

[0136] As a preferred embodiment of the present application, the identification terminal is a smart phone.

[0137] As a preferred embodiment of the present application, the identification terminal is a smart watch.

[0138] As a preferred embodiment of the present application, the identification terminal is any device having a development board, a processing chip, a gyroscope sensor and a wireless communication module, and the identification terminal is provided with a shell for carrying and fixing.

[0139] As a preferred embodiment of the present application, the identification terminal is a laptop computer.

[0140] As a preferred embodiment of the present application, the identification terminal is a camera with an Internet of Things function, wherein the camera is used to be carried or fixed in a certain environment, the number of cameras is not limited, and the cameras are connected to each other through the Internet of Things;

[0141] When this embodiment is implemented, for the same target, the same location area or the unified instruction, at least two cameras are associated and coordinated. Specifically, one camera recognizes one of the key information data, and the subsequent camera performs supplementary recognition of the remaining key information data:

[0142] For the same person target, location area or accepted task instruction, at least one of the camera terminals records the person's walking picture and exchanges data. When the person enters the monitoring area of ​​another camera, the camera confirms the information and captures the motion video, and processes the motion trajectory and the person's posture recognition data during the trajectory in the background.

[0143] As another preferred embodiment of the present application, the identification terminal is a combination of one or more of the following terminal devices, including a smart phone; a smart watch; any device having a development board, a processing chip, a gyroscope sensor and a wireless communication module; a laptop computer; a camera with an Internet of Things function;

[0144] In a specific embodiment of the present application, the terminal interconnection communication module is used to realize data communication with the cloud and data communication between multiple terminals via wireless means;

[0145] The data feedback module is used to obtain the data and terminal information and obtain the corresponding data feedback through the cloud;

[0146] It should be noted that the data obtained includes computing logic, computing consensus mechanism, trained computing model, training logic, and training method.

[0147] The consensus communication module is used to issue consensus logic instructions for unified execution to the terminal devices within the range according to the data, type, location, environment and current status information of the connected terminal devices;

[0148] The terminal feedback confirmation module is used to obtain the terminal location information, wearing location information, information disclosed by the wearing user, environmental weather information and terminal density information through the feedback information obtained from the terminal feedback and network communication;

[0149] The local operation module is used to call the local computing component to perform local operation, storage and training of data according to the consensus logic instructions;

[0150] The terminal independent training module is used to obtain the training model, training method and training data adapted to the data received by the current terminal through the local operation module and data communication, and perform independent operation training on such data through the computing component;

[0151] The data interaction and synchronization module is used to synchronize and interact the local information and operation information of the terminal according to logic. At the same time, it is also used to communicate data among multiple compatible terminals within the range and share data and computing power;

[0152] A data classification module is used to classify the acquired data, and store and operate the classified data as a data set;

[0153] The posture evaluation module is used to calculate and analyze the posture of the current terminal wearer and communicate the evaluation logic and evaluation results.

[0154] In this embodiment, it is not difficult to see that the identification terminal interacts with the cloud before use to determine the crowd information, environmental information and location information of the identification terminal, and when data is modified through the cloud, multi-terminal verification and confirmation are required;

[0155] The identification terminal is worn on the body and obtains the current environmental data through hardware components, network communications and installed sensors. The environmental data includes weather data, location data, information on the number of terminals within the location range, and the current location information of the terminal worn on the human body;

[0156] Subsequently, the recognition logic and recognition template are determined according to the current environment, and the data is annotated with location, environment, and weather, and network data training and local data training are performed. In addition, the local data cannot be used directly, and is trained for a certain period of time through preset logic or preset model. At the same time, within the same range, the device transmits data through wireless communication protocols, obtains local training data of other devices, and performs synchronization and supplementation;

[0157] When pedestrians wear identification terminals, the corresponding environmental data is periodically fed back. The movement trajectory is not recorded, but the area environment of the movement can be recorded.

[0158] In addition, in this embodiment, multiple identification terminals are interconnected through a wireless communication protocol, and request command communication is performed on a terminal with a network communication function. This terminal with a network communication function is used to perform network data communication, data request and data transmission. Multiple identification terminals without a network are connected through any compatible protocol such as Bluetooth, WiFi LAN, Zigbee, etc.

[0159] Embodiment 3

[0160] like Figure 3 As shown, based on the above embodiment, step 1 specifically further includes the following steps:

[0161] Step 1.1: Download Market-1501, Dukemtmc-Reid, and Cuhk03 datasets.

[0162] Step 1.2: Image preprocessing:

[0163] For each image, cropping and scaling are required to unify the image size to 256*128 pixels. At the same time, image normalization is required to map the pixel values ​​to the range of [0,1].

[0164] Step 1.3: Data Augmentation:

[0165] To increase the diversity of training data, each image requires random horizontal flipping, random cropping, and other data augmentation operations.

[0166] Step 2 specifically also includes the following steps:

[0167] Step 2.1: Feature Extraction:

[0168] Use Resnet-50 as the backbone network to extract the feature vector of pedestrian images;

[0169] Step 2.2: Feature vector fusion:

[0170] The feature vectors are fused using the feature distance fusion method to obtain a more comprehensive and accurate representation of pedestrian features;

[0171] Step 2.3: Person Re-identification:

[0172] The pedestrian feature vector is classified using the SoftMax classifier to obtain the identity of the pedestrian.

[0173] Step 3 specifically also includes the following steps:

[0174] Step 3.1: Feature Extraction:

[0175] Input pedestrian images into the pre-trained CNN model to extract feature representations of the images;

[0176] Step 3.2: Posture information evaluation:

[0177] Evaluate pedestrian posture information;

[0178] Step 3.3: Iterative optimization:

[0179] The pose estimation problem is converted into an optimization problem and the pose parameters of the pedestrian are optimized iteratively.

[0180] Step 4 specifically also includes the following steps:

[0181] Step 4.1: Pose Estimation Model Training:

[0182] The MPII Human Pose Dataset, an annotated human pose dataset, is used for model training.

[0183] Step 4.2: Pose Estimation Model Evaluation:

[0184] After training, the model is evaluated using a test dataset, the predictions are compared with the annotated data, and the error is calculated;

[0185] Step 4.3: Application of pose estimation model:

[0186] Apply the trained pose estimation model to person re-identification and construct pedestrian feature representation by predicting pedestrian pose information;

[0187] Step 4.4: Integration of feature representations:

[0188] The feature representations obtained from person re-identification and pose estimation models are combined into a single feature representation. This is achieved by concatenating the feature vectors obtained from the two models into a single vector. The resulting feature representation is used for matching and retrieval tasks.

[0189] Step 4.5: Evaluation:

[0190] The performance of the proposed method is evaluated on benchmark datasets using standard evaluation metrics such as Rank-1 accuracy, Mean Average Precision (MAP), and False Positive Rate (FPR).

[0191] As another preferred embodiment of the present application, the algorithm and training model are transplanted to the recognition terminal for use and communication connection. The specific implementation steps are as follows:

[0192] (1) Hardware design:

[0193] The Jetson Nano development board is used as the hardware platform, equipped with NVIDIA's GPU accelerator and 4GB of memory, which can meet the operating requirements of the pedestrian re-identification algorithm.

[0194] Install JetPack SDK: JetPack SDK is a software development kit provided by NVIDIA for the Jetson platform, which includes various commonly used software tools and library files. First, you need to install JetPack SDK on Jetson Nano using the method provided in the official documentation.

[0195] Install the deep learning framework: Install the TensorFlow deep learning framework. Use the package management tool provided by JetPack SDK to install it.

[0196] (2) Software design: In order to realize the transplantation of algorithms and models, TensorRT is used for model optimization, and the trained model is converted into a TensorRT engine that can run efficiently on Jetson Nano. At the same time, this project also uses OpenCV for image processing and visualization of recognition results, realizing real-time monitoring and management of pedestrian re-identification.

[0197] First, convert the person re-identification model into TensorRT format to improve the inference efficiency on embedded systems such as Jetson Nano. The process is mainly divided into the following steps:

[0198] Use deep learning frameworks such as TensorFlow to train the model; convert the trained model to ONNX format; use the TensorRT API to convert the ONNX model to the TensorRT engine; optimize the parameters of the TensorRT engine to improve the inference speed and accuracy.

[0199] Connect the camera and test: Connect the camera to the Jetson Nano and test it to ensure that the person re-identification algorithm can read the camera image normally and output the correct person re-identification results.

[0200] (3) Product production: Based on the above hardware and software design, a portable pedestrian re-identification device is produced. The device adopts a compact design and is easy to carry, which makes it convenient for users to perform pedestrian re-identification tasks in different scenarios.

[0201] (4) Testing and verification: Conduct multiple sets of tests to verify the performance of pedestrian re-identification in different scenarios.

[0202] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A pedestrian re-identification method, characterized in that: The method employs an identification terminal and an identification system for the identification terminal; The method comprises the following steps: Step 1: Prepare training dataset: Prepare the Market-1501, Dukemtmc-Reid and Cuhk03 datasets and train them. Before training, preprocess these datasets, including image preprocessing and data augmentation. Step 2: Establish a pose estimation algorithm and perform training and optimization: During training, select and tune hyperparameters including learning rate, loss function, and regularization term, and use cross-validation for model selection and performance evaluation; Step 3: Implement the pose estimation algorithm: Use deep learning posture estimation algorithms to build posture estimation models and match the appearance of pedestrians in different postures to estimate the posture information of pedestrians; Step 4: Adjust and optimize the pose estimation module: Training, evaluating and optimizing the posture evaluation model, inputting the feature representation extracted by the CNN model and the current posture parameters into the posture estimation module, obtaining new posture parameters, and updating the optimized objective function, iterating the process until convergence, that is, the optimized objective function reaches a stable state; Step 5: Get the recognition results and transfer the model as needed: Transplant the algorithm and training model into the recognition terminal, and realize the recognition and communication transmission of the device.

2. The pedestrian re-identification method according to claim 1, characterized in that: The identification terminal is worn on any suitable part of the human body through a strap, and the identification terminal is any terminal device with computing and communication capabilities; The identification terminal includes a terminal interconnection communication module, a data feedback module, a consensus communication module, a terminal feedback confirmation module, a local operation module, a terminal independent training module, a data interaction synchronization module, a posture evaluation module and a data classification module, wherein: The terminal interconnection communication module, the data feedback module and the consensus communication module are all connected to the terminal feedback confirmation module and the local operation module, the terminal feedback confirmation module and the local operation module are all connected to the terminal independent training module and the data interaction synchronization module, and the terminal independent training module and the data interaction synchronization module are all connected to the posture assessment module and the data classification module.

3. The pedestrian re-identification method according to claim 2, characterized in that: The terminal interconnection communication module is used to realize data communication with the cloud and data communication between multiple terminals via wireless mode; The data feedback module is used to obtain the data and terminal information and obtain the corresponding data feedback through the cloud; The consensus communication module is used to issue consensus logic instructions for unified execution to the terminal devices within the range according to the data, type, location, environment and current status information of the connected terminal devices; The terminal feedback confirmation module is used to obtain the terminal location information, wearing location information, information disclosed by the wearing user, environmental weather information and terminal density information through the feedback information obtained from the terminal feedback and network communication; The local operation module is used to call the local computing component to perform local operation, storage and training of data according to the consensus logic instructions; The terminal independent training module is used to obtain the training model, training method and training data adapted to the data received by the current terminal through the local operation module and data communication, and perform independent operation training on such data through the computing component; The data interaction and synchronization module is used to synchronize and interact the local information and operation information of the terminal according to logic. At the same time, it is also used to communicate data among multiple compatible terminals within the range and share data and computing power; A data classification module is used to classify the acquired data, and store and operate the classified data as a data set; The posture evaluation module is used to calculate and analyze the posture of the current terminal wearer and communicate the evaluation logic and evaluation results.

4. The pedestrian re-identification method according to claim 1 or 3, characterized in that: Step 1 specifically also includes the following steps: Step 1.1: Download Market-1501, Dukemtmc-Reid and Cuhk03 datasets; Step 1.2: Image preprocessing: For each image, cropping and scaling are required to unify the image size to 256*128 pixels. At the same time, image normalization is required to map the pixel values ​​to the range of [0,1]. Step 1.3: Data Augmentation: To increase the diversity of training data, each image requires random horizontal flipping, random cropping, and other data augmentation operations.

5. The pedestrian re-identification method according to claim 1 or 3, characterized in that: Step 2 specifically also includes the following steps: Step 2.1: Feature Extraction: Use Resnet-50 as the backbone network to extract the feature vector of pedestrian images; Step 2.2: Feature vector fusion: The feature vectors are fused using the feature distance fusion method to obtain a more comprehensive and accurate representation of pedestrian features; Step 2.3: Person Re-identification: The pedestrian feature vector is classified using the SoftMax classifier to obtain the identity of the pedestrian.

6. The pedestrian re-identification method according to claim 1 or 3, characterized in that: Step 3 specifically also includes the following steps: Step 3.1: Feature Extraction: Input pedestrian images into the pre-trained CNN model to extract feature representations of the images; Step 3.2: Posture information evaluation: Evaluate pedestrian posture information; Step 3.3: Iterative optimization: The pose estimation problem is converted into an optimization problem and the pose parameters of the pedestrian are optimized iteratively.

7. The pedestrian re-identification method according to claim 1 or 3, characterized in that: Step 4 specifically also includes the following steps: Step 4.1: Pose Estimation Model Training: Use the annotated human pose dataset MPII Human Pose Dataset for model training; Step 4.2: Pose Estimation Model Evaluation: After training, the model is evaluated using a test dataset, the predictions are compared with the annotated data, and the error is calculated; Step 4.3: Application of pose estimation model: Apply the trained pose estimation model to person re-identification and construct pedestrian feature representation by predicting pedestrian pose information; Step 4.4: Integration of feature representations: The feature representations obtained from the person re-identification and pose estimation models are combined into a single feature representation. This is achieved by concatenating the feature vectors obtained from the two models into a single vector. The resulting feature representation is used for matching and retrieval tasks. Step 4.5: Evaluation: The performance of the proposed method is evaluated on benchmark datasets using standard evaluation metrics.

Images