Lightweight cross-modal pedestrian re-identification method for edge computing device, storage medium and device

Abstract

The application discloses a lightweight cross-modal pedestrian re-identification method for edge computing devices, a storage medium and equipment, comprising a lightweight residual module fusing a depth separable convolution, a channel attention mechanism and an instance-batch normalization, which is replaced into a high-level stage of a standard residual network, while maintaining feature expression capability, and greatly reducing model computing power requirements; at the same time, the method constructs a joint loss function, fusing a standard triplet loss, a quadruplet loss, a weighted regularization triplet loss and an optimal transport alignment loss, and strengthens the discriminability and modal invariance of cross-modal features; and a preheating-cosine annealing learning rate scheduling strategy is combined for model optimization. The method of the application significantly reduces the model complexity under the premise of ensuring the recognition accuracy, can be efficiently deployed on edge devices, and meets the real-time end-side inference requirements in large-scale urban monitoring scenarios.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, specifically relating to a lightweight cross-modal pedestrian re-identification method, storage medium, and device for edge computing devices. Background Technology

[0002] Cross-modal person re-identification (Cross-modal Re-ID) is a key technology in computer vision. Its core objective is to perform cross-scene and cross-modal identity matching and verification of specific pedestrians in different camera scenarios. It is widely used in practical scenarios such as intelligent surveillance, public safety, and intelligent transportation. With the rapid development of edge computing technology, deploying cross-modal person re-identification models on edge devices to achieve real-time, localized cross-modal person identification has become a current research hotspot and practical need in the field.

[0003] In the development of cross-modal person re-identification technology, deep learning-based models have become mainstream. Currently, cross-modal person re-identification models generally use Residual Networks (ResNet) as the backbone network. Although ResNet has strong feature extraction capabilities, it suffers from inherent drawbacks such as large parameter count, high computational complexity, and slow inference speed, making it difficult to meet the deployment requirements of edge devices for lightweight, low-power, and real-time inference. To achieve model lightweighting, existing research has attempted to introduce various lightweight architectures. Among them, OSNet (Omni-Scale Network), with its multi-scale feature fusion and efficient convolutional design, can significantly reduce the number of parameters and computational cost while maintaining strong feature representation capabilities. It has been validated and applied in single-modal person re-identification tasks.

[0004] However, existing technologies lack cross-modal person re-identification models using OSNet as the backbone network, and there is also a lack of mature solutions that combine the lightweight advantages of OSNet with ResNet to improve traditional ResNet-based cross-modal models. Therefore, how to leverage OSNet to lightweightly improve ResNet-based cross-modal person re-identification models, while ensuring cross-modal recognition accuracy, and addressing the technical pain points of excessive model size and inability to be deployed on edge devices, has become a key issue that urgently needs to be addressed in this field. Summary of the Invention

[0005] This invention addresses the problems of excessively large size and insufficient recognition accuracy in existing cross-modal pedestrian re-identification models. It provides a lightweight cross-modal pedestrian re-identification method, storage medium, and device for edge computing. This includes a lightweight residual module that integrates depthwise separable convolution, channel attention mechanism, and instance-batch normalization, replacing the high-level stage of the standard residual network. This significantly reduces the model's computational requirements while maintaining feature representation capabilities. Furthermore, the method constructs a joint loss function that integrates standard triplet loss, quadruple loss, weighted regularized triplet loss, and optimal transmission alignment loss, enhancing the discriminative power and modality invariance of cross-modal features. A preheating-cosine annealing learning rate scheduling strategy is also incorporated for model optimization. This invention significantly reduces model complexity while maintaining recognition accuracy, enabling efficient deployment on edge devices and meeting the real-time edge inference requirements of large-scale urban surveillance scenarios.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a lightweight cross-modal pedestrian re-identification method for edge computing devices, a basic cross-modal pedestrian re-identification model based on a two-stream network structure, and the construction of a lightweight feature extraction backbone network and a joint loss function;

[0007] The lightweight feature extraction backbone network: In the deep stages of the residual network, the original standard residual modules are replaced with lightweight residual modules; the lightweight residual modules are specifically: cascaded 1×1 pointwise convolutions and grouped 3×3 depthwise separable convolutions are used to replace the 3×3 convolutions in the standard residual modules; the number of channels in the intermediate layers of the modules is set to the maximum value between one-quarter of the number of input channels and the preset minimum number of channels; instance-batch normalization is used, and a channel attention module is embedded after the residual connections;

[0008] The joint loss function includes standard triplet loss, quadruplet loss, weighted regularized triplet loss, and optimal transmission loss for aligning the distributions of visible and infrared modal features. The joint loss function is optimized using a preheating-cosine annealing learning rate scheduling strategy.

[0009] As an improvement to the present invention, the depthwise separable convolutional layer in the lightweight residual module is decomposed into two independent operational layers: the depthwise convolutional kernel dimension is... The pointwise convolution kernel dimension is ;in, This represents the convolution operation. The convolution kernel spatial size, The number of input feature channels, The number of output feature channels; for a height of Width is Input tensor Its theoretical calculation cost is The number of parameters is .

[0010] As another improvement of the present invention, the standard triplet loss in the joint loss function is specifically as follows:

[0011] )

[0012] in, This is an example of an anchor. This is a positive example. It is a negative example. It is a constant greater than 0; (This refers to the Euclidean distance or cosine distance metric function between feature vectors, used to quantify the similarity differences between different samples in the embedding space;)

[0013] The specific quaternion loss is as follows:

[0014]

[0015] in, These are negative examples of samples that are not of the same category. It is the boundary of the triplet loss. It is the boundary of the quadruple extension term;

[0016] The weighted regularized triplet loss in the joint loss function is specifically as follows:

[0017]

[0018] in, For sample weights, For the triplet loss boundary, For regularization weights, The square of the L2 norm; , , These represent the feature vectors extracted from the anchor sample, positive sample, and negative sample by the backbone network, respectively.

[0019] The optimal transmission alignment loss is specifically:

[0020] in, It is the first One visible light mode feature, It is the first One infrared modal feature, Representation of features and The Euclidean distance between them; These are the soft-matching weight matrix elements obtained through the Sinkhorn algorithm, used to describe the optimal transmission relationship between visible light samples and infrared samples; It is a mask for samples of the same type; it takes 1 if they share the same identity, and 0 otherwise. It is a mask for outlier samples, set to 1 for different identities and 0 otherwise; It is the weight coefficient of the outlier separation term, used to balance the alignment of similar categories and the repulsion of outliers; This indicates the batch size during model training.

[0021] As another improvement of this invention, a preheating-cosine annealing learning rate scheduling strategy is adopted to optimize the joint loss function, specifically as follows:

[0022]

[0023] in, It is the learning rate at step t. It is the base learning rate. This represents the total number of warm-up steps, where t is the current training step count. It is the minimum learning rate. It is the maximum learning rate. It is the number of iterations in the current cycle. It is the total number of steps in a cosine annealing cycle.

[0024] As a further improvement of the present invention, in the basic cross-modal pedestrian re-identification model of the dual-stream network structure, the SYSU-MM01 and / or LLCM datasets are used for training, and an adversarial occlusion data augmentation strategy is adopted for data augmentation. The adversarial occlusion data augmentation strategy specifically involves: randomly selecting a rectangular region in the image with a preset probability, filling the region with the pixel mean, and controlling the augmentation intensity by adjusting the lower limit, upper limit, and aspect ratio range of the erased region.

[0025] To achieve the above objectives, the present invention also adopts the following technical solution: a non-transitory machine-readable storage medium storing executable code thereon, wherein when the executable code is executed by the processor of an electronic device, the processor executes the lightweight cross-modal pedestrian re-identification method for edge computing devices as described above.

[0026] To achieve the above objectives, the present invention also adopts the following technical solution: a computer device, comprising:

[0027] Memory, on which executable code is stored;

[0028] A processor is configured to execute the executable code, causing the computer device to perform the operation of the lightweight cross-modal pedestrian re-identification method for edge computing devices.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The present invention proposes a simple and effective cross-modal person re-identification model based on OSNet lightweight improvement. This model integrates the lightweight structure of OSNet with the ResNet backbone network, so that the cross-modal person re-identification model can significantly reduce the number of parameters and computational complexity while maintaining high-precision feature expression.

[0031] (2) The method of this invention applies the OSNet lightweight architecture, which is designed for pedestrian re-identification, to cross-modal pedestrian re-identification tasks for the first time, and achieves lightweight improvement of the traditional ResNet-based model, filling the gap in the existing technology of lacking a lightweight, edge-deployable cross-modal pedestrian re-identification scheme.

[0032] (3) The lightweight cross-modal pedestrian re-identification model proposed in this invention has a simple structure and can realize real-time feature extraction and deployment on edge devices. It effectively solves the technical pain point that traditional large models cannot be implemented on large-scale urban surveillance cameras. It is accurate in calculation and significantly improves inference effect. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of the cross-modal pedestrian re-identification network model containing a lightweight feature extraction backbone network of the present invention. Detailed Implementation

[0034] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0035] Example 1

[0036] Existing cross-modal pedestrian re-identification models mostly employ heavily parameterized backbone networks, resulting in a large number of parameters and high computational complexity, making them difficult to deploy on edge devices with limited computing power. Therefore, this invention proposes a lightweight cross-modal pedestrian re-identification method for edge computing devices. Based on a dual-stream network structure, a lightweight feature extraction backbone network and joint loss function are constructed, significantly reducing model complexity while ensuring recognition accuracy. This method can be efficiently deployed on edge devices, meeting the real-time edge inference requirements of large-scale urban surveillance scenarios.

[0037] This invention first establishes a basic cross-modal person re-identification model. This basic model employs a two-stream network structure, a novel network model for visible-infrared person re-identification (VIReID) tasks, primarily addressing the modal gap between visible light (VIS) and infrared (IR) images. In this embodiment, a cross-modal person re-identification model with ResNet-50 as its backbone is selected as the basic model to be improved. This model learns more informative feature representations by generating diverse embedding vectors in the embedding space, effectively reducing the modal differences between visible light and infrared images. The publicly available cross-modal person re-identification datasets SYSU-MM01 and LLCM are used as training and testing data. Data augmentation is performed using adversarial occlusion: image regions are randomly selected with a set probability and filled with the mean to simulate occlusion and sensor noise in real-world scenarios. The shape of the occluded region is adjusted by parameters such as the upper and lower limits of the erased area and the aspect ratio range, improving the model's robustness in complex scenarios.

[0038] A lightweight feature extraction backbone network is constructed based on a fundamental cross-modal pedestrian re-identification model with a two-stream network structure. For example... Figure 1 As shown, the lightweight feature extraction backbone network: In the deep stages of the residual network, the original standard residual modules are replaced with lightweight residual modules, that is, cascaded 1×1 pointwise convolutions and grouped 3×3 depthwise separable convolutions are used to replace the 3×3 convolutions in the standard residual modules; the number of channels in the intermediate layers of the modules is set to the maximum value between one-quarter of the number of input channels and a preset minimum number of channels; instance-batch normalization is used, and a channel attention module is embedded after the residual connections. The lightweight residual module is named lightBottleneck.

[0039] The basic idea behind the depthwise separable convolution used in the lightweight residual module is to decompose a convolutional layer into two independent operational layers: a depthwise convolutional kernel with a dimension of 1. The pointwise convolution kernel dimension is ;in, This represents the convolution operation. The convolution kernel spatial size, The number of input feature channels, This represents the number of output feature channels. For a height of... Width is Input tensor Its theoretical calculation cost is The number of parameters is .

[0040] In this embodiment, the lightweight residual module is only applied to the third and fourth residual convolutional stages in the residual network of the basic cross-modal person re-identification model. All standard residual modules in the third residual convolutional stage are replaced with lightweight residual modules, and the first two standard residual modules in the fourth residual convolutional stage are replaced with lightweight residual modules. The last residual module remains the original standard residual module. The ResNet-50 network structure with the lightweight residual module is shown in the table below:

[0041] Table 1

[0042]

[0043] As shown in the table above, the overall topology of the lightweight backbone network is as follows: The input image is first downsampled through a 7×7 convolutional layer with a stride of 2 and a 3×3 max-pooling layer; then it passes through four residual stages. The first stage contains three standard Bottleneck modules, with an output feature map size of 56×56; the second stage contains four standard Bottleneck modules, with an output feature map size of 28×28; the third stage is the deep semantic feature extraction region, containing six consecutive lightweight residual modules (light Bottleneck), with an output feature map size of 14×14; the fourth stage further abstracts global features, replacing the first two modules with light Bottleneck, and the last module retains the original standard Bottleneck structure to maintain high-level semantic integrity, with an output feature map size of 7×7. At the end of the network, a global average pooling layer is used to compress the spatial dimension to 1×1, and then mapped to a fixed-dimensional identity feature vector through a fully connected layer.

[0044] Based on the aforementioned cross-modal pedestrian re-identification model that includes a lightweight feature extraction backbone network, a joint loss function is constructed, and a preheating-cosine annealing learning rate scheduling strategy is adopted to optimize the joint loss function.

[0045] The joint loss function includes standard triplet loss, quadruplet loss, weighted regularized triplet loss, and optimal transmission loss for aligning visible and infrared modal feature distributions, enhancing the discriminative power and modal invariance of cross-modal features and improving model recognition accuracy.

[0046] The standard triplet loss is used to bring anchor samples closer to positive samples and widen the distance between anchor samples and negative samples, specifically as follows:

[0047] )

[0048] in, This is an example of an anchor. It is a positive example, and They are samples of the same category. It is a negative example, and Samples that are not of the same category It is a constant greater than 0.

[0049] The quadruple loss introduces a second negative sample on top of the binary loss to further constrain intra-class compactness and inter-class separation, specifically:

[0050]

[0051] in, This is an example of an anchor; It is a positive example, and They are samples of the same category; It is a negative example, and Samples that are not of the same category They are not samples of the same category. It is the boundary margin of the triplet loss, which controls the minimum distance difference between features of the same identity and features of different identities. It is the boundary of the quadruple extension term, controlling the minimum difference between the distance between people of the same identity and the distance between people of different identities.

[0052] Weighted regularized triplet loss introduces sample weights and L2 regularization to improve the ability to mine hard samples, specifically:

[0053]

[0054] in, For sample weights, for Boundaries control the minimum distance difference between individuals with the same identity and those with different identity characteristics. It is the regularization weight. It is the square of the L2 norm.

[0055] The optimal transmission alignment loss is solved by using the Sinkhorn algorithm to calculate the soft matching weights between visible light and infrared samples, thereby narrowing the distance between cross-modal samples of the same identity and widening the distance between cross-modal samples of different identities. Specifically:

[0056]

[0057] in, It is the first One visible light mode feature, It is the first One infrared modal feature, This represents the Euclidean distance between the two. It is a soft matching weight obtained by the Sinkhorn algorithm, used to describe the optimal transmission relationship between visible light samples and infrared samples; It is a mask for similar samples, when and If they belong to the same identity, use 1; otherwise, use 0. It is a mask for outlier samples, when and If they do not belong to the same identity, use 1; otherwise, use 0. This is the weight of the outlier separation term, used to balance alignment between similar and outlier categories; in the current implementation, it is set to 0.5. This loss improves the model's cross-modal feature alignment capability by narrowing the distance between cross-modal same-identity samples and increasing the distance between cross-modal different-identity samples.

[0058] The training process employs a preheating-cosine annealing learning rate scheduling strategy to optimize the joint loss function. During the preheating phase, the learning rate linearly increases from zero to a base learning rate to stabilize the training process. After preheating, the learning rate periodically decays according to a cosine function to help the model escape local optima and converge to a better solution. The expression is:

[0059]

[0060] in, It is the learning rate at step t. It is the base learning rate. This represents the total number of warm-up steps, where t is the current training step count. It is the minimum learning rate. It is the maximum learning rate. It is the number of iterations in the current cycle. It is the total number of steps in a cosine annealing cycle.

[0061] Finally, the trained lightweight cross-modal pedestrian re-identification model is converted into an inference format adapted to embedded platforms and deployed to a Jetson Nano edge computing device. Pedestrian image feature extraction and matching are performed directly at the camera end, with only the feature vectors uploaded to the server, avoiding the bandwidth and privacy pressures of raw video transmission. This enables real-time cross-modal pedestrian re-identification inference on the edge in large-scale urban surveillance scenarios.

[0062] Test case

[0063] To verify the effectiveness and advancement of the method of this invention, this test case selects representative algorithms in the field of cross-modal person re-identification in recent years as the comparison baseline, such as the early classic methods TONE, AGW, and DDAG, and recent mainstream methods such as PMT, PIC, PMDA, WF-CAMReViT, TMD, DARD, and DEEN. All comparative test experiments were conducted under a unified hardware platform and data partitioning protocol: the standard full search / indoor search partitioning strategy of the SYSU-MM01 and LLCM datasets was adopted. During the training phase, the AdamW optimizer was used, and the multivariate loss function was jointly optimized by combining the Warmup and cosine annealing learning rate scheduling strategies. During the inference phase, the feature vectors of the query set and the image library set were directly extracted, a cross-modal cosine similarity matrix was constructed and sorted for retrieval, and finally, the four core evaluation indicators of Rank-1, Rank-10, Rank-20, and mAP were statistically analyzed. After multiple rounds of training and parameter tuning, Tables 2 and 3 are obtained below. Among them, Table 2 shows the performance comparison values ​​of the method of this invention and the mainstream methods on the SYSU-MM01 dataset for the above four evaluation indicators:

[0064] Table 2

[0065]

[0066] As shown in Table 2, the method of this invention (Ours) achieved a Rank-1 score of 70.65% and an mAP of 64.72% in the SYSU-MM01 full search mode. Compared with earlier methods, such as TONE, BDTR, and AlignGAN, this invention achieves a significant lead in all metrics, especially with a Rank-1 improvement of over 20 percentage points. Compared with advanced methods in recent years, such as PMT, SPOT, and DARD, this invention still maintains a leading advantage in the Rank-1 metric, improving by 1.32% compared to DARD and 3.12% compared to PMT. Furthermore, the Rank-10 and Rank-20 scores reach 95.14% and 98.69%, respectively. Therefore, Table 2 demonstrates that the lightweight module and joint loss function proposed in this invention have significant advantages in cross-modal feature alignment and discriminative mining, with both recall and precision reaching advanced levels in the industry.

[0067] Table 3 below shows the performance comparison values ​​of the method of this invention and various mainstream methods on the LLCM dataset for the above four evaluation metrics:

[0068] Table 3

[0069]

[0070] As shown in Table 3, on the LLCM dataset with complex lighting conditions and greater modal differences, the method of this invention performs particularly well in the VIS→IR retrieval direction, achieving a Rank-1 of 65.6%, which is 3.1% higher than the current mainstream baseline DEEN's 62.5%. In the IR→VIS direction, the Rank-1 of this invention is 53.3%, which is similar to DEEN, but the R10 and R20 indicators, at 82.0% and 88.7% respectively, remain at a high level. The combined bidirectional retrieval results show that the method of this invention exhibits excellent modality-invariant generalization ability in complex scenarios such as low lighting and cross-viewpoint conditions, verifying that the lightweight architecture does not sacrifice the model's feature representation ability in extreme environments.

[0071] The test results above demonstrate that this invention achieves excellent cross-modal retrieval performance on the publicly available datasets SYSU-MM01 and LLCM. While significantly compressing the number of parameters to approximately 15M and meeting the low-power deployment requirements at the edge, the model achieves advanced levels of cross-modal retrieval accuracy, effectively addressing the technical pain points of traditional large models, such as high computational requirements and difficulty in deployment on city-level monitoring terminals. Higher values ​​for each evaluation metric indicate stronger identity matching accuracy and feature discrimination capabilities. Through lightweight residual module reconstruction and multi-objective joint loss optimization, experiments on the SYSU-MM01 and LLCM datasets fully validate its technical feasibility and robustness in engineering deployment scenarios.

[0072] In summary, the method of this invention adopts the OSNet lightweight architecture to improve the traditional cross-modal person re-identification model based on ResNet, and combines multiple loss functions for joint optimization to improve the model's recognition accuracy, ultimately achieving efficient deployment and real-time inference of the cross-modal person re-identification model on edge devices.

[0073] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A lightweight cross-modal pedestrian re-identification method for edge computing devices, characterized in that... A basic cross-modal pedestrian re-identification model based on a two-stream network structure is constructed, and a lightweight feature extraction backbone network and joint loss function are built. The lightweight feature extraction backbone network: In the deep stages of the residual network, the original standard residual modules are replaced with lightweight residual modules; the lightweight residual modules are specifically: cascaded 1×1 pointwise convolutions and grouped 3×3 depthwise separable convolutions are used to replace the 3×3 convolutions in the standard residual modules; the number of channels in the intermediate layers of the modules is set to the maximum value between one-quarter of the number of input channels and the preset minimum number of channels; instance-batch normalization is used, and a channel attention module is embedded after the residual connections; The joint loss function includes standard triplet loss, quadruplet loss, weighted regularized triplet loss, and optimal transmission loss for aligning visible and infrared modal feature distributions; the joint loss function is optimized using a preheating-cosine annealing learning rate scheduling strategy.

2. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 1, characterized in that: In the lightweight feature extraction backbone network, the lightweight residual module is only applied to the third and fourth residual convolutions in the residual network of the basic cross-modal pedestrian re-identification model. The third residual convolution consists entirely of lightweight residual modules, and the first two standard residual modules of the fourth residual convolution are lightweight residual modules, while the last residual module remains the original standard residual module.

3. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 2, characterized in that: The depthwise separable convolutional layer in the lightweight residual module is decomposed into two independent operational layers: the depthwise convolutional kernel dimension is... The pointwise convolution kernel dimension is ;in, This represents the convolution operation. The convolution kernel spatial size, The number of input feature channels, This determines the number of output feature channels.

4. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 1, characterized in that: The standard triplet loss in the joint loss function is specifically: ) ； in, This is an example of an anchor. This is a positive example. It is a negative example. It is a constant greater than 0; The Euclidean distance metric function represents the distance between feature vectors; The specific quaternion loss is as follows: ； in, These are negative examples of samples that are not of the same category. It is the boundary of the triplet loss. It is the boundary of the quadruple extension.

5. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 4, characterized in that: The weighted regularized triplet loss in the joint loss function is specifically as follows: ； in, For sample weights, For the triplet loss boundary, For regularization weights, The square of the L2 norm; , , These represent the feature vectors extracted from the anchor sample, positive sample, and negative sample by the backbone network, respectively. The optimal transmission alignment loss is specifically: ； in, It is the first One visible light mode feature, It is the first One infrared modal feature, Representation of features and The Euclidean distance between them; These are elements of the soft matching weight matrix; It is a mask for similar samples. It is a mask for outlier samples; These are the weighting coefficients of the heterogeneous separation terms; This indicates the batch size during model training.

6. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 5, characterized in that: The joint loss function is optimized using a preheating-cosine annealing learning rate scheduling strategy, specifically as follows: ； in, It is the learning rate at step t. It is the base learning rate. This represents the total number of warm-up steps, where t is the current training step count. It is the minimum learning rate. It is the maximum learning rate. It is the number of iterations in the current cycle. It is the total number of steps in a cosine annealing cycle.

7. The lightweight cross-modal pedestrian re-identification method for edge computing devices as described in claim 1, characterized in that: In the basic cross-modal pedestrian re-identification model of the dual-stream network structure, the SYSU-MM01 and / or LLCM datasets are used for training, and an adversarial occlusion data augmentation strategy is adopted for data augmentation. The adversarial occlusion data augmentation strategy is as follows: a rectangular region in the image is randomly selected with a preset probability, and the region is filled with the pixel mean. The augmentation intensity is controlled by adjusting the lower limit, upper limit and aspect ratio range of the erased region.

8. A non-transitory machine-readable storage medium, characterized in that: It stores executable code that, when executed by the processor of an electronic device, causes the processor to perform a lightweight cross-modal pedestrian re-identification method for edge computing devices as described in any one of claims 1-7.

9. A computer device, characterized in that: include: Memory, on which executable code is stored; A processor is configured to execute the executable code, causing the computer device to perform the operation of the lightweight cross-modal pedestrian re-identification method for edge computing devices as described in any one of claims 1-8.

Images