A pedestrian detection and re-identification method, system and terminal based on deep learning

Abstract

The application belongs to the technical field of deep learning, and particularly relates to a pedestrian detection and re-identification method, system and terminal based on deep learning, which comprises the following steps: S1, data preprocessing: arranging and dividing pedestrian images according to IDs into a training set and a test set, and the images in the sets are not repeated; S2, constructing a pedestrian re-identification model: using a deep learning framework PyTorch to build an L2GNet model, which is composed of 17 convolution layers in total, and each convolution uses ReLU as an activation function; S3, model training: training the L2GNet on the training set, and iteratively updating the parameters of the network according to the back propagation principle; S4, inputting the test set data into the trained model to perform a detection and re-identification task, calculating a pedestrian re-identification performance index, and saving the queried pedestrian images. The application builds a pedestrian detection and re-identification model, and the trained model can be used for pedestrian detection and re-identification tasks.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, and in particular relates to a method, system and terminal for pedestrian detection and re-identification based on deep learning. Background Technology

[0002] The core of building an intelligent surveillance system is the automatic processing of image data captured by cameras, with pedestrians being the most important focus in these images. Therefore, the detection, identification, and tracking of pedestrians in real-world scenarios has significant application value. Currently, widely used facial recognition technology has limitations, requiring a high degree of cooperation from personnel: it needs to meet many conditions, such as viewing angle, lighting, occlusion, and shooting distance. Clearly, most surveillance equipment in real-world scenarios cannot acquire such high-quality image data. Deep learning, due to its superior image feature representation capabilities, is widely used in image processing tasks.

[0003] Pedestrian detection and re-identification technology utilizes computer vision to extract and compare features of pedestrians appearing in images or videos, thereby enabling the tracking and identification of the same pedestrian across different time periods and under different monitoring conditions. Its task is to match one or a set of detected images with images in a database to identify the identity of the detected images (set). It is widely used in video surveillance, including pedestrian search, camera tracking, and behavior analysis. However, it inevitably faces the challenge of diverse real-world scenarios. Since the pedestrian to be queried may appear under multiple different cameras, different devices have different resolutions and camera sizes, and the pedestrian's posture, viewpoint, and background constantly change in different monitoring videos, all of these significantly affect image quality and thus the accuracy of the model. With the advent of deep learning technology, compared to traditional methods, deep learning-based pedestrian detection and re-identification algorithms have achieved significant performance improvements.

[0004] Currently, pedestrian detection and re-identification can be studied from two aspects: feature representation and similarity measurement.

[0005] However, existing person re-identification algorithms still lag significantly behind practical applications. Despite advancements, many methods exhibit suboptimal performance in challenging scenarios. For instance, traditional research utilizes techniques like pictorial structure and segmentation to process pedestrian images, then matches them using low-level visual features and fits the difference distribution between samples using a Gaussian model. However, variations in pose and occlusion lead to numerous matching errors. While some research focuses on pose estimation methods, these often require additional models to provide prior information, limiting their efficiency. Furthermore, with the increasing richness and complexity of person re-identification datasets, relying solely on global features is insufficient to achieve satisfactory performance, limiting accuracy.

[0006] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: the current pedestrian re-identification algorithm only uses global features or additional prior models, which cannot balance performance and efficiency, and the accuracy of pedestrian detection and re-identification still needs to be improved. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method, system, and terminal for pedestrian detection and re-identification based on deep learning.

[0008] This invention is implemented as follows: a deep learning-based pedestrian detection and re-identification method, comprising:

[0009] S1, Data preprocessing: Pedestrian images are sorted by ID and divided into training set and test set, and the images in the sets are not repeated;

[0010] S2, Constructing a pedestrian re-identification model: The L2GNet model is built using the deep learning framework PyTorch. The model consists of 17 convolutional layers, with each convolution using ReLU as the activation function.

[0011] S3, Model Training: Train L2GNet on the training set and iteratively update the network's parameters according to the backpropagation principle;

[0012] S4. Input the test set data into the trained model to perform detection and re-identification tasks, calculate the pedestrian re-identification performance index, and save the retrieved pedestrian images.

[0013] Furthermore, the 17 convolutional layers of the pedestrian re-identification model in S2 are divided into 4 convolutional blocks: the first layer, the hidden layer, and the last layer.

[0014] The first layer contains three convolutional layers and one ReLU activation function, and the number of channels is increased from 3 to 64.

[0015] The hidden layer contains 15 convolutional layers, each followed by a ReLU activation function, and the residual structure design keeps the number of channels constant.

[0016] The final layer is divided into a global branch and a local branch, which consists of three convolutional layers and a ReLU activation function, restoring the number of channels from 64 to 3. The local branch adds an extra 1×1 convolution, reducing the number of channels in the local feature map after horizontal pooling from 2048 dimensions to 128 dimensions to improve computational efficiency.

[0017] Furthermore, S3 trains the L2GNet model on the training set, performs gradient optimization on the model using the Adam optimization algorithm, and updates the parameters using a loss function combining the joint cross-entropy loss function and the hard-sample triplet loss function.

[0018]

[0019] L th =(max p∈A d a,p -min n∈B d a,n +α) + ;

[0020] L = L softmax +L th ;

[0021] Where n represents the number of training samples in each batch, and α represents the margin between samples, which is set according to actual needs.

[0022] Furthermore, the same prefix network ResNet50 is used to extract front-end features. The output is then divided into two branches: one is to directly apply global pooling to the feature map, thereby extracting global features into a global branch; the other is to extract local features of each row through horizontal pooling, and then apply 1×1 convolution to reduce the number of channels. The final loss is calculated by the hard sample triplet loss.

[0023] Another object of the present invention is to provide a deep learning-based pedestrian detection and re-identification system that applies the aforementioned deep learning-based pedestrian detection and re-identification method. The deep learning-based pedestrian detection and re-identification system includes:

[0024] The data preprocessing module is used to divide pedestrian images into training and test sets according to their IDs, ensuring that the images in each set are unique.

[0025] The model building module is used to build L2GNet models using the deep learning framework PyTorch.

[0026] The model training module trains L2GNet on the training set and iteratively updates the network's parameters based on the backpropagation principle.

[0027] The recognition module is used to input test set data into the trained model for detection and re-recognition tasks, calculate pedestrian re-recognition performance indicators, and save the retrieved pedestrian images.

[0028] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the deep learning-based pedestrian detection and re-identification method.

[0029] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based pedestrian detection and re-identification method.

[0030] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned deep learning-based pedestrian detection and re-identification system.

[0031] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0032] First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0033] This invention improves the loss function by introducing a loss function that combines softmax loss and hard sample triplet loss, which exhibits a competitive effect in measuring the similarity between images.

[0034] This invention adds additional local branches to the high-level feature representation in deep learning models and learns partial / region aggregated features, making the model more robust to problems such as misalignment.

[0035] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0036] This invention uses the PyTorch framework to build an L2GNet model, which is based on and improved from the ResNet50 network to make it suitable for pedestrian detection and re-identification tasks. During training, the average precision value is used as the model training evaluation metric, and the best-performing model is saved in each iteration. The saved model is then used to evaluate the detection and re-identification performance on a test set. This invention achieves high-precision pedestrian detection and re-identification performance on public datasets and achieves excellent results on multiple evaluation metrics. Specifically, when compared with methods from the past 5 years, this invention exhibits better performance in most cases, further validating the effectiveness of the improved method.

[0037] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0038] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0039] In recent years, with the continuous development of the national economy and technology and the continuous improvement of public living standards, social security issues have received increasing attention. The rising demands for public safety have led to an explosive growth in the demand for intelligent monitoring applications, and the deployment of numerous surveillance cameras has provided a guarantee for maintaining public safety. However, it is not difficult to see that the video data generated in real time by these surveillance devices is enormous and is growing exponentially over time. Therefore, when abnormal events occur and it is necessary to obtain information about people's movements or identify individuals from the surveillance footage, it often requires a significant amount of manpower and resources to manually collect and retrieve this massive amount of data, resulting in poor timeliness.

[0040] The pedestrian detection and re-identification method provided by this invention employs advanced deep learning technology. Based on the improved ResNet50 model, it can track and re-identify pedestrians across different cameras, achieving more accurate identification and location. This will facilitate timely identification of target individuals. It is anticipated that after widespread application in the public safety field, this technology will significantly reduce crime rates, enhance public safety, and provide law enforcement agencies with more effective tools. Simultaneously, the application of this invention's technical solution will also improve the intelligence of surveillance cameras, alleviate the pressure of manual monitoring, and reduce policing costs. It is expected that in the coming years, the technical solution provided by this invention will generate significant economic benefits while also having a positive social impact on the overall safety environment of cities.

[0041] (2) Whether the technical solution of the present invention overcomes technical bias:

[0042] The technical solution of this invention effectively overcomes potential technical biases in existing pedestrian detection and re-identification technologies by employing advanced loss metric and pre-trained weight methods. These biases may stem from differences in data distribution, domain shifts between datasets, etc., leading to misleading results from traditional methods.

[0043] To overcome technical bias, the technical solution provided by this invention uses pre-trained weight parameters as initialization guidance during model training, and preserves local detail features and employs advanced loss measurement methods during re-recognition, ensuring excellent recognition accuracy even with limited training samples. For example, traditional methods suffer from decreased recognition accuracy in certain regions due to sample scarcity, especially in environments protecting data privacy and individual portrait rights, where data scarcity is a difficult problem to solve. The technical solution provided by this invention effectively overcomes this bias, making its practical application and economic benefits more significant.

[0044] Fourth, in the embodiments of the present invention, each step of the deep learning-based pedestrian detection and re-identification method has brought about significant technological advancements:

[0045] S1, Data Preprocessing: The preprocessing step greatly improves the efficiency and quality of model training. By arranging pedestrian images according to their IDs and dividing them into training and test sets, we ensure that the images in the sets are not duplicated, thereby avoiding overfitting and enabling the model to better generalize to new, unseen data.

[0046] S2, Building a Person Re-identification Model: A deep learning model, L2GNet, specifically tailored for person re-identification, was constructed. This model utilizes the PyTorch deep learning framework and consists of 17 convolutional layers, each using ReLU as the activation function, resulting in a model architecture with good performance and adaptability. In particular, the residual structure design and 1×1 convolutions effectively maintain the consistency of the number of channels while reducing the dimensionality of local feature maps, thereby improving computational efficiency.

[0047] S3, Model Training: During the training of the L2GNet model on the training set, the Adam optimization algorithm and the joint cross-entropy loss function and hard sample triplet loss function were used. This enabled the network to better identify the characteristics of different pedestrians during the learning process and to effectively handle hard samples, thereby improving the model's performance. Furthermore, by gradually reducing the learning rate, the model converged better, thus improving its stability and prediction accuracy.

[0048] S4, Detection and Re-identification: Input the test set data into the trained model to perform detection and re-identification tasks, calculate the pedestrian re-identification performance index, and save the retrieved pedestrian images. This enables the model to perform effective pedestrian detection and re-identification in practical applications, thereby realizing the practicality of the model and providing an important foundation for subsequent pedestrian tracking and analysis.

[0049] In summary, this deep learning-based pedestrian detection and re-identification method, through its carefully designed model structure and training strategy, achieves efficient and accurate pedestrian detection and re-identification, demonstrating high practical value and broad application prospects. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating the steps of the deep learning-based pedestrian detection and re-identification method provided in this embodiment of the invention.

[0052] Figure 2 This is a flowchart of a deep learning-based pedestrian detection and re-identification method provided in an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram of the network structure provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the test results for the image similarity detection method provided in the embodiments of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] To address the problems existing in the prior art, this invention provides a method, system, and terminal for pedestrian detection and re-identification based on deep learning. The invention will be described in detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown, the pedestrian detection and re-identification method based on deep learning provided in this embodiment of the invention includes the following steps:

[0058] S1, Data preprocessing: Pedestrian images are sorted by ID and divided into training set and test set, and the images in the sets are not repeated;

[0059] S2, Constructing a pedestrian re-identification model: The L2GNet model is built using the deep learning framework PyTorch. The model consists of 17 convolutional layers, with each convolution using ReLU as the activation function.

[0060] The pedestrian re-identification model in S2 has 17 convolutional layers divided into 4 convolutional blocks: the first layer, the hidden layer, and the last layer.

[0061] The first layer contains three convolutional layers and one ReLU activation function, and the number of channels is increased from 3 to 64.

[0062] The hidden layer contains 15 convolutional layers, each followed by a ReLU activation function, and the residual structure design keeps the number of channels constant.

[0063] The final layer is divided into a global branch and a local branch, each consisting of three convolutional layers and a ReLU activation function, reducing the number of channels from 64 to 3. The local branch adds an extra 1×1 convolution, reducing the number of channels in the horizontally pooled local feature map from 2048 dimensions to 128 dimensions to improve computational efficiency.

[0064] S3, Model Training: Train L2GNet on the training set and iteratively update the network's parameters according to the backpropagation principle;

[0065] S3 trains an L2GNet model on the training set, uses the Adam optimization algorithm for gradient optimization, and updates parameters using a loss function combining the joint cross-entropy loss function and the hard-sample triplet loss function.

[0066]

[0067] L th =(max p∈A d a,p -min n∈B d a,n +α) + ;

[0068] L = L softmax +L th ;

[0069] Wherein, the input image x i It was identified as having the tag y i The predicted probability is encoded using the softmax function, given by p(y i |x i ) represents; n represents the number of training samples in each batch, a represents the anchor image, p represents the positive sample image, n represents the negative sample image, d represents the distance between features of samples; α represents the margin between samples, set to 0.3;

[0070] The network was trained using this joint loss function, and the model parameters were optimized using the Adam optimizer. The initial learning rate was set to 0.0002, and the network was trained for a total of 300 epochs. At half the time, the learning rate was reduced to one-tenth of its original value.

[0071] S4. Input the test set data into the trained model to perform detection and re-identification tasks, calculate the pedestrian re-identification performance index, and save the retrieved pedestrian images.

[0072] like Figure 2 The diagram shown is a flowchart of the pedestrian detection and re-identification method provided in an embodiment of the present invention.

[0073] like Figure 3 The diagram shown is a schematic of the network structure provided in an embodiment of the present invention. The same prefix network (ResNet50) is used to extract the front-end features. The output is divided into two branches: one is to directly apply global pooling to the feature map, thereby extracting global features into a global branch; the other is to extract local features of each row through horizontal pooling, and then apply 1×1 convolution to reduce the number of channels. The final loss is calculated by the hard sample triplet loss.

[0074] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of a deep learning-based pedestrian detection and re-identification method.

[0075] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of a deep learning-based pedestrian detection and re-identification method.

[0076] An application embodiment of the present invention provides an information data processing terminal, which is used to implement a pedestrian detection and re-identification system based on deep learning.

[0077] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.

[0078] like Figure 4 The table shows the test results of the image similarity detection method provided by the embodiments of the present invention. To better illustrate the advantages and contributions of the present invention in solving pedestrian detection and re-identification tasks, it is compared with some other existing methods to evaluate their performance and effectiveness. As shown in Table 1, the model of the embodiments of the present invention using the joint loss function of softmax and hard sample triples for similarity measurement shows higher performance in multiple evaluation metrics than other single loss functions and their combinations. As shown in Table 2, after the introduction of the new joint loss function, the pedestrian re-identification model of the embodiments of the present invention, which uses local branches and global branches for joint training, shows higher performance in multiple evaluation metrics than ResNet50 which only uses the global branch, especially with a 7% higher mean accuracy (mAP). Meanwhile, the performance of other different methods on the Market1501 dataset is also compared. As shown in Table 3, the embodiments of the present invention achieve high-precision pedestrian detection and re-identification performance on the Market1501 dataset and achieve good results on multiple evaluation metrics, indicating that the proposed method exhibits excellent competitiveness under test conditions. Furthermore, the performance of this invention was evaluated on the challenging MSMT17 dataset, which consists of 180 hours of video containing 4101 pedestrian identities. Each pedestrian typically has images from 15 camera perspectives, exhibiting complex lighting variations. This dataset is the closest to real-world conditions in terms of scale, occlusion, lighting, and perspective. As shown in Table 4, the embodiments of this invention still demonstrate highly competitive accuracy compared to other methods.

[0079] Table 1. Performance comparison of the pedestrian similarity measurement method of the present invention with various other measurement methods.

[0080] loss function Rank-1 Rank-5 Rank-10 mAP Softmax loss 0.86 0.94 0.97 0.70 Center loss 0.78 0.90 0.93 0.58 Softmax+Center 0.78 0.90 0.93 0.62 Center+TriHard 0.71 0.87 0.91 0.52 TriHard+Softmax 0.87 0.95 0.97 0.71

[0081] Table 2 Performance comparison between the pedestrian re-identification method of the present invention and ResNet50 using only global branches.

[0082] Loss function Rank-1 Rank-5 Rank-10 mAP ResNet50 0.87 0.95 0.97 0.71 L2GNet 0.91 0.97 0.98 0.78

[0083] Table 3. Performance comparison of the pedestrian re-identification method of this invention with other methods on the Market1501 dataset.

[0084]

[0085]

[0086] Table 4. Performance comparison of the pedestrian re-identification method of this invention with other methods on the MSMT17 dataset.

[0087] Methods Rank-1 Rank-5 Rank-10 mAP PDC 0.58 0.73 0.79 0.30 Joint 0.60 0.76 0.81 0.31 GLAD 0.61 0.77 0.81 0.34 L2GNet 0.64 0.79 0.84 0.37

[0088] Example 1: Application of Deep Learning-Based Pedestrian Detection and Re-identification in a Shopping Mall Surveillance System

[0089] Data preprocessing: Collect pedestrian images from the shopping mall's surveillance system, extract the pedestrians' faces using facial recognition technology, and sort them according to each pedestrian's ID, dividing them into training and test sets.

[0090] Build a pedestrian re-identification model: Using the PyTorch framework, build the L2GNet model according to the given architecture.

[0091] Model training: The model was trained using the Adam optimizer and joint loss functions (cross-entropy loss and hard triplet loss). During training, the learning rate was set to 0.0002, and then reduced to one-tenth of its original value halfway through the training.

[0092] Detection and Re-identification: After the model is trained, the data from the test set is input into the model to perform pedestrian detection and re-identification. The retrieved pedestrian images are also saved for subsequent pedestrian tracking and analysis.

[0093] Example 2: Application of Deep Learning-Based Pedestrian Detection and Re-identification in Intelligent Transportation Systems

[0094] Data preprocessing: Pedestrian images are collected from the traffic monitoring system, and full-body images of pedestrians are extracted using computer vision technology. The images are then sorted according to each pedestrian's ID and divided into training and test sets.

[0095] Build a pedestrian re-identification model: Using the PyTorch framework, build the L2GNet model according to the given architecture.

[0096] Model training: The model was trained using the Adam optimizer and joint loss functions (cross-entropy loss and hard triplet loss). During training, the learning rate was set to 0.0002, and then reduced to one-tenth of its original value halfway through the training.

[0097] Detection and Re-identification: After the model is trained, the data from the test set is input into the model for pedestrian detection and re-identification. The retrieved pedestrian images are also saved for subsequent pedestrian tracking and analysis. In intelligent transportation systems, this can be used for pedestrian safety, traffic planning, and many other aspects.

[0098] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A pedestrian detection and re-identification method based on deep learning, characterized in that, The L2GNet model, built using the deep learning framework PyTorch, is used for pedestrian re-identification. This L2GNet model has 17 convolutional layers and uses ReLU as the activation function in each convolutional layer, thereby achieving efficient and accurate performance in pedestrian detection and re-identification tasks. include: S1, Data preprocessing: Pedestrian images are sorted by ID and divided into training set and test set, and the images in the sets are not repeated; S2, Constructing a pedestrian re-identification model: The L2GNet model is built using the deep learning framework PyTorch. The model consists of 17 convolutional layers, with each convolution using ReLU as the activation function. S3, Model Training: Train L2GNet on the training set and iteratively update the network's parameters according to the backpropagation principle; S4. Input the test set data into the trained model to perform detection and re-identification tasks, calculate the pedestrian re-identification performance index, and save the queried pedestrian images. The pedestrian re-identification model in S2 has 17 convolutional layers divided into 4 convolutional blocks: the first layer, the hidden layer, and the last layer. The first layer contains three convolutional layers and one ReLU activation function, and the number of channels is increased from 3 to 64. The hidden layer contains 15 convolutional layers, each followed by a ReLU activation function, and the residual structure design keeps the number of channels constant. The final layer is divided into a global branch and a local branch, which consists of three convolutional layers and a ReLU activation function, restoring the number of channels from 64 to 3. The local branch adds an extra 1×1 convolution, reducing the number of channels in the local feature map after horizontal pooling from 2048 dimensions to 128 dimensions to improve computational efficiency.

2. The pedestrian detection and re-identification method based on deep learning as described in claim 1, characterized in that, S3 trains an L2GNet model on the training set, uses the Adam optimization algorithm for gradient optimization, and updates parameters using a loss function combining the joint cross-entropy loss function and the hard-sample triplet loss function. ； ； ； Where n represents the number of training samples in each batch, and α represents the margin between samples, which is set according to actual needs.

3. The pedestrian detection and re-identification method based on deep learning as described in claim 1, characterized in that, The same prefix network ResNet50 is used to extract front-end features. The output is then divided into two branches: one is to directly apply global pooling to the feature map, thereby extracting global features into a global branch; the other is to extract local features of each row through horizontal pooling, and then apply 1×1 convolution to reduce the number of channels. The final loss is calculated by the hard sample triplet loss.

4. A deep learning-based pedestrian detection and re-identification system applying the deep learning-based pedestrian detection and re-identification method as described in any one of claims 1 to 3, characterized in that, Deep learning-based pedestrian detection and re-identification systems include: The data preprocessing module is used to divide pedestrian images into training and test sets according to their IDs, ensuring that the images in each set are unique. The model building module is used to build L2GNet models using the deep learning framework PyTorch. The model training module trains L2GNet on the training set and iteratively updates the network's parameters based on the backpropagation principle. The recognition module is used to input test set data into the trained model for detection and re-recognition tasks, calculate pedestrian re-recognition performance indicators, and save the retrieved pedestrian images.

5. A computer device comprising a memory and a processor, the memory storing a computer program, wherein when the computer program is executed by the processor, the processor performs the steps of the deep learning-based pedestrian detection and re-identification method as described in any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based pedestrian detection and re-identification method as described in any one of claims 1 to 3.

7. An information data processing terminal, the information data processing terminal being used to implement the deep learning-based pedestrian detection and re-identification system as described in claim 4.

Images