Lightweight pedestrian detection method, system and device based on improved SSD algorithm and medium

By improving the SSD algorithm and combining it with MobileNetV3-Small and attention mechanisms, a lightweight pedestrian detection model was constructed, which solved the accuracy and real-time problems of pedestrian detection in complex scenarios and enabled efficient deployment on resource-constrained devices.

CN118968615AInactive Publication Date: 2024-11-15NORTH CHINA INST OF AEROSPACE ENG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410985711.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing pedestrian detection methods suffer from low detection accuracy and poor real-time performance when dealing with complex scenarios such as occlusion, scale changes, and small pedestrian targets, and are difficult to deploy on resource-constrained devices.

Method used

We employ an improved SSD algorithm, combined with the MobileNetV3-Small base network, SE and CA attention mechanisms, and construct a lightweight pedestrian detection model through dilated convolution, cross-layer skip connections, and customized anchor box optimization, thereby enhancing feature extraction and detection capabilities.

Benefits of technology

It improves the accuracy and real-time performance of pedestrian detection and enables efficient deployment on resource-constrained devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968615A_ABST
    Figure CN118968615A_ABST
Patent Text Reader

Abstract

The invention discloses a lightweight pedestrian detection method, system and device based on an improved SSD algorithm and a medium, and relates to the technical field of computer vision. The method comprises the following steps: detecting an obtained target input image by using a lightweight pedestrian detection model to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed by taking MobileNetV3-Small as a basic network, combining SE and CA attention mechanisms and improving an SSD (Solid State Disk) algorithm; according to the improved SSD algorithm, feature information is further extracted by adding an additional convolutional layer behind a basic convolutional network. According to the method, the detection precision and the real-time performance can be improved, and efficient deployment on resource-constrained equipment is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a lightweight pedestrian detection method, system, device and medium based on an improved SSD algorithm. Background Art

[0002] Pedestrian detection technology is an important research direction in the field of computer vision and is widely used in intelligent transportation systems, intelligent security monitoring, intelligent robots, and pedestrian flow monitoring. Traditional pedestrian detection methods rely on manually designed features and complex image processing techniques, but these methods have problems with low detection accuracy and poor real-time performance when dealing with complex scenes such as occlusion, scale changes, and small pedestrian targets. With the development of deep learning technology, deep learning-based methods have made significant progress in pedestrian detection, but existing methods often require a lot of computing resources and are difficult to deploy on resource-constrained devices. Summary of the invention

[0003] The purpose of the present invention is to provide a lightweight pedestrian detection method, system, device and medium based on an improved SSD algorithm, which can improve detection accuracy and real-time performance and achieve efficient deployment on resource-constrained devices.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A lightweight pedestrian detection method based on an improved SSD algorithm includes:

[0006] Get the target input image;

[0007] The target input image is detected using a lightweight pedestrian detection model to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed using MobileNetV3-Small as the basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding an additional convolutional layer after the basic convolutional network.

[0008] Optionally, the method for constructing the lightweight pedestrian detection model is:

[0009] S1. Backbone network construction: MobileNetV3-Small is selected as the basic network, and an improved MNetV3-Smallv2 network model is constructed by introducing SE and CA attention mechanisms;

[0010] S2, image resizing: resize the input image from the original 224×224 to 300×300;

[0011] S3. Application of dilated convolution module: Use the dilated convolution module to replace the first ordinary convolution layer in the basic network; the dilated convolution module is an expanded convolution layer with a dilation rate of (1, 2);

[0012] S4, CA attention mechanism integration: Introduce the CoordinateAttention mechanism in the Bneck block;

[0013] S5. Cross-layer skip connection design: According to the feature size and channel of the Bneck layer, a lightweight cross-layer skip module is designed, and the ECA attention mechanism is introduced;

[0014] S6. Anchor frame customization optimization: The anchor frame is redesigned based on the Anchor mechanism, three aspect ratios are set, and the minimum prior frame size on the feature map is adjusted;

[0015] S7, lightweight design of backend network: Improve the backend network of the SSD algorithm, use 1×1 convolution to increase the dimension, 3×3 channel-by-channel convolution for feature learning, and 1×1 point-by-point convolution for dimensionality reduction and information fusion;

[0016] S8, ECA attention mechanism and H-Swish activation function application: Introduce ECA attention mechanism and H-Swish activation function in the additional layer;

[0017] S9, lightweight pedestrian detection model construction: replace the original backbone network VGG16, cut the MNetV3-Smallv2 network, retain the key feature extraction module, and build a lightweight pedestrian detection model;

[0018] S10. Experimental verification and performance evaluation: Experimental verification is performed on the Caltech pedestrian dataset to evaluate the model’s performance indicators such as mAP and FPS.

[0019] Optionally, the dilated convolution module in S3 is applied as follows: in the first 3×3 convolution layer in the Dilated Convolution block, the padding is 1, the convolution step is 2, and the dilation rate is 1; in the second 3×3 convolution layer, the padding is 2, the convolution step is 1, and the dilation rate is 2.

[0020] Optionally, the method for integrating the CA attention mechanism in S4 is: generating position weights by calculating the global average pooling of the horizontal and vertical coordinates of the input feature map, and then combining these weights with the original feature map to enhance the feature representation.

[0021] Optionally, the process of lightweight design of the back-end network in S7 is: first use a 1×1 convolution layer in the additional layer to increase the dimension, then use a 3×3 channel-by-channel convolution to replace the ordinary convolution layer, and introduce an ECA attention mechanism and an H-Swish activation function in the second additional layer, and finally use a 1×1 point-by-point convolution to reduce the dimension.

[0022] The present invention also provides a lightweight pedestrian detection system based on an improved SSD algorithm, comprising:

[0023] An image acquisition module, used to acquire a target input image;

[0024] The image detection module is used to detect the target input image using a lightweight pedestrian detection model to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed with MobileNetV3-Small as the basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding an additional convolutional layer after the basic convolutional network.

[0025] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-mentioned lightweight pedestrian detection method based on the improved SSD algorithm.

[0026] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the lightweight pedestrian detection method based on the improved SSD algorithm as described above.

[0027] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0028] The present invention discloses a lightweight pedestrian detection method, system, device and medium based on an improved SSD algorithm. The method comprises using a lightweight pedestrian detection model to detect an acquired target input image to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed using MobileNetV3-Small as a basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding an additional convolutional layer after the basic convolutional network. The present invention can improve detection accuracy and real-time performance, and achieve efficient deployment on resource-constrained devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 This is a schematic diagram of the design of the MobileNetV3-Small network architecture in this embodiment;

[0031] Figure 2 This is a schematic diagram of the design of the MNetV3-Smallv2 network architecture in this embodiment;

[0032] Figure 3 This is the MNetV3-Smallv2 backbone network structure diagram in this embodiment;

[0033] Figure 4 is a structural diagram of the Bneck layer of the lightweight cross-layer skip connection module in this embodiment;

[0034] Figure 5 This is a CA-Bneck structure diagram in this embodiment;

[0035] Figure 6 Schematic diagram of the algorithm flow of the CA attention mechanism in this embodiment;

[0036] Figure 7 This is the flow chart of the SE attention mechanism in this embodiment;

[0037] Figure 8 This is the structure diagram of the ECA attention mechanism in this embodiment;

[0038] Fig. 9 It is the SKIP1 NET cross-layer connection module in this embodiment;

[0039] Fig.10 It is the SKIP2 NET cross-layer connection module in this embodiment;

[0040] Fig.11 It is the SKIP3 NET cross-layer connection module in this embodiment;

[0041] Fig.12 : is a structural diagram of an additional convolutional layer in this embodiment;

[0042] Fig.13 is a structural diagram of the improved additional convolutional layer in this embodiment;

[0043] Fig.14 This is a schematic diagram of the MNetV3-Smallv2-SSD model structure design in this embodiment;

[0044] Fig.15 Schematic diagram of the dilated convolution module in this embodiment;

[0045] Fig.16 Schematic diagram of the lightweight pedestrian detection method in this embodiment. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The purpose of the present invention is to provide a lightweight pedestrian detection method, system, device and medium based on an improved SSD algorithm, which can improve detection accuracy and real-time performance and achieve efficient deployment on resource-constrained devices.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] The present invention provides a lightweight pedestrian detection method based on an improved SSD algorithm, comprising:

[0050] Step 100: Obtain a target input image;

[0051] Step 200: Use a lightweight pedestrian detection model to detect the target input image to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed with MobileNetV3-Small as the basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding an additional convolutional layer after the basic convolutional network.

[0052] As a further implementation method, the method for constructing a lightweight pedestrian detection model is:

[0053] S1. Backbone network construction: MobileNetV3-Small is selected as the basic network, and an improved MNetV3-Smallv2 network model is proposed. The feature expression capability is enhanced by introducing SE and CA attention mechanisms.

[0054] S2. Image resizing: The input image size is resized from the original 224×224 to 300×300 to capture more visual information.

[0055] S3. Application of dilated convolution module: Use an expanded convolution layer with a dilation rate of (1, 2) to replace the first ordinary convolution layer to expand the network's receptive field and enhance the perception of multi-scale information.

[0056] S4, CA attention mechanism integration: Introduce the CoordinateAttention mechanism in the Bneck block to focus on the key local areas in the input feature map and improve the accuracy of feature capture.

[0057] S5. Cross-layer skip connection design: According to the feature size and channel of the Bneck layer, a lightweight cross-layer skip module is designed, and the ECA attention mechanism is introduced to enrich feature information and improve learning ability.

[0058] S6. Anchor frame customization optimization: The anchor frame is redesigned based on the Anchor mechanism, using aspect ratios (0.2, 0.4, 0.6), and the minimum prior frame size on the feature map is adjusted to accommodate the annotation of pedestrian targets.

[0059] S7. Lightweight design of back-end network: Improve the back-end network of the SSD algorithm, use 1×1 convolution to increase the dimension, 3×3 channel-by-channel convolution for feature learning, and 1×1 point-by-point convolution for dimensionality reduction and information fusion.

[0060] S8, ECA attention mechanism and H-Swish activation function application: Introduce ECA attention mechanism and H-Swish activation function in the additional layer to effectively capture the correlation between channels and improve model performance.

[0061] S9. Construction of lightweight pedestrian detection model MNetV3-Smallv2-SSD: Replace the original backbone network VGG16, trim the MNetV3-Smallv2 network, retain the key feature extraction module, and build a lightweight model suitable for pedestrian detection.

[0062] S10. Experimental verification and performance evaluation: Experimental verification is conducted on the Caltech pedestrian dataset to evaluate the model’s performance indicators such as mAP and FPS to ensure the effectiveness of the model in practical applications.

[0063] The application of the dilated convolution module in step S3 is as follows: the padding value in the first 3×3 convolution layer in the dilated convolution block is 1, the convolution step size is 2, and the dilation rate is 1; the padding value in the second 3×3 convolution layer is 2, the convolution step size is 1, and the dilation rate is 2.

[0064] The CA attention mechanism integration in step S5 is as follows: by calculating the global average pooling of the horizontal and vertical coordinates of the input feature map, position weights are generated, and then these weights are combined with the original feature map to enhance the feature representation.

[0065] The lightweight design of the back-end network in step S7: The specific improvement method is to first use a 1×1 convolution layer in the additional layer to increase the dimension, then use a 3×3 channel-by-channel convolution to replace the ordinary convolution layer, and introduce the ECA attention mechanism and H-Swish activation function in the second additional layer, and finally use a 1×1 point-by-point convolution to reduce the dimension.

[0066] Based on the above technical solution, the following is provided: Figure 1-Figure 16 The embodiment shown.

[0067] 1. Innovative improvements to the backbone network

[0068] MNetV3-Smallv2 model

[0069] Based on the MobileNetV3-Small model (see the attached network architecture Figure 1 ) It consists of 4 ordinary convolutional layers, 11 Bneck layers, and an average pooling layer, and proposes an improved MNetV3-Smallv2 network model (see the network architecture Figure 2 ), SE attention mechanism is referenced in all Bneck layers except the second and third Bneck layers, then CA attention mechanism is referenced in the second and third Bneck layers. It should be noted that unlike the original paper of MobileNetV3, the 1*1 convolution for latitude ascending is removed in the first Bneck layer in the improved model of this paper because the number of channels is not expanded when passing through the first Bneck layer, and the SE attention mechanism is removed in the first convolution layer after the last Bneck block to reduce excessive dependence on the MNetV3-Smallv2 backend network.

[0070] Image size

[0071] The input image size of MNetV3-Smallv2 is adjusted from the original input image size of 224*224 to 300*300.

[0072] Application of the dilated convolution module

[0073] The hole convolution module is used to replace the first layer of ordinary convolution in the original network to deepen the network structure and expand the receptive field, thereby enhancing the perception of global and multi-scale information.

[0074] First, replace the first ordinary convolution layer of the original network with a dilated convolution layer with a dilation rate of (1, 2) to increase the receptive field size. The actual convolution kernel size of the dilated convolution is calculated as follows:

[0075] k′=k+(k-1)(r-1)

[0076] Where k is the convolution kernel size and r is the dilation rate. The current receptive field of the dilated convolution is calculated as follows:

[0077] RF i+1 =RF i +(k'-1)×S i

[0078] Among them, RF i+1 Represents the receptive field of the current layer, RF i It represents the receptive field of the previous layer. And when designing the dilated convolution rate, the HDC principle should be followed, as shown in the following formula:

[0079] M i =max[M i+1 -2r i ,M i+1 -2(M i+1 -r i ),r i ]

[0080] Among them, M i Represents the distance between non-zero elements in the i-th layer, r i represents the expansion coefficient of the i-th layer. It should be noted that no matter which layer its M i It should always be less than or equal to the size of the k convolution kernel. This will deepen the network depth, expand the convolution receptive field, and increase the context information of the feature map without changing the original feature map output size and number of channels. The structure of the dilated convolution module is shown in the attached figure. Fig.15 As shown in the figure, in the Dilated Convolution block, the padding value in the first 3*3 convolution layer is 1, the convolution step size is 2, and the dilation rate is 1. The padding value in the second 3*3 convolution layer is 2, the convolution step size is 1, and the dilation rate is 2.

[0081] Integration of CA attention mechanism

[0082] The advantage of the CA attention mechanism is that it comprehensively considers the channel and direction-related position information, and has sufficient flexibility and lightness, and can be easily embedded in the core module of a lightweight network. This technology introduces the CoordinateAttention (CA) mechanism in the Bneck block (the structure diagram is shown in the attached figure). Figure 5 The structural process is shown in the attached Figure 6 As shown in Figure 2, it focuses on the key local areas in the input feature map to improve the accuracy and computational efficiency of feature capture.

[0083] Innovative design of cross-layer skip connections

[0084] According to the size and channel of the input and output feature layers of the Backeck layers numbered 1, 2, 4, 7, and 9 in the original MobileNetV3-Small network, three lightweight cross-layer skip modules are designed (see Appendix Fig. 9 , 10 , 11) and introduced the ECA attention mechanism, namely SKIP1 Net, SKIP2 Net and SKIP3Net, to enrich the feature information, improve the learning ability of the network and achieve lightweight design.

[0085] ECA attention mechanism (structure diagram as attached Figure 8 As shown in the figure, it is a lightweight and efficient plug-and-play attention mechanism based on the SE attention mechanism (the algorithm flowchart is shown in the attached figure). Figure 7 The SE attention mechanism is an improved version of the SE block. Given the input features, the SE block first uses global average pooling for each channel separately, then uses two fully connected (FC) layers with nonlinearity, and then uses a Sigmoid function to generate channel weights. However, dimensionality reduction has side effects on channel attention prediction and destroys the direct correspondence between channels and their weights. Therefore, the ECA mechanism is selected to use a one-dimensional convolution method to avoid the impact of dimensionality reduction on the data. It first squeezes the feature map. This step is to convert the feature map from a size of (N, C, H, W) to (N, C, 1, 1) through global average pooling, so that the fusion of global context information is achieved. This step is the same as the SE attention mechanism. Then calculate the size of the adaptive convolution kernel, as shown in the following formula. Among them, C is the number of input channels, b = 1, γ = 2, and a one-dimensional convolution is used to calculate the channel weight. Finally, the Sigmoid activation function is used to map the weight between (0-1). Finally, the reshaped weight value is multiplied by the original feature map to obtain the feature map under different weights.

[0086]

[0087] SKIP1 NET cross-layer connection module: The output feature map of the empty convolution module passes through the cross-layer skip connection module of the first layer, and then passes through the 3*3 channel-by-channel convolution layer of the skip module to output a feature map with the same output size and number of channels as the current Bneck layer, and the two feature maps are spliced ​​to form a feature map containing more detailed information.

[0088] SKIP2 NET cross-layer connection module: Since the output feature map size and number of channels of the previous Bneck layer are different from the output feature map of the current Bneck layer, the cross-layer skip connection module of this layer first uses channel-by-channel convolution to keep the input size of the previous layer consistent with the input size of the current layer, and then unifies the number of channels through point-by-point convolution, and finally connects the feature maps. Here, ordinary convolution is not used for downsampling, but channel-by-channel convolution is used, which has the advantage of reducing the amount of calculation and the number of parameters.

[0089] SKIP3 NET cross-layer connection module: The size of the feature map output by the current layer is consistent with that of the previous layer, with only the number of channels being different.

[0090] 2. Customized Optimization of Anchor Frame Design

[0091] According to the particularity of pedestrian targets, we redesigned the three aspect ratios of anchor boxes based on the Anchor mechanism: 0.2, 0.4, and 0.6. In addition, the minimum size of the prior box on the 38*38 feature map is set to 20, and the minimum size of the prior box on the 19*19 feature map is set to 50, while the rest remain unchanged, making it more suitable for marking pedestrian targets.

[0092] 3. Lightweighting and performance improvement of backend network

[0093] Lightweight design of backend network

[0094] The back-end network of the SSD algorithm is improved, and the second and third additional layers are improved. First, 1*1 convolution is used to increase the dimension, 3*3 channel-by-channel convolution is used for feature learning, and finally 1*1 point-by-point convolution is used to reduce the dimension and blend information.

[0095] The original SSD target detection model consists of a VGG16 backbone network with additional layers, prediction box generation layers, category prediction layers, and bounding box regression layers. SSD (Single Shot Multibox Detector) adds additional convolutional layers after the basic convolutional network to extract richer feature information at different scales, thereby achieving efficient detection of multi-scale targets. The design of this set of additional convolutional layers allows SSD to complete the target detection task in a single forward propagation. At the same time, through the introduction of multiple prediction heads and anchor boxes, it can effectively capture targets of different scales and aspect ratios, thereby realizing an end-to-end efficient detection model. The additional layer has 4 additional convolutional layer structures as shown in the attached figure. Fig.12 shown.

[0096] These four additional convolutional layers are composed of 4 groups of 1*1 convolutional layers and 3*3 ordinary convolutional layers, respectively. They extract multi-scale feature maps of sizes 10*10, 5*5, 3*3 and 1*1 for the generation of prior frames and direct classification and regression. According to the composition characteristics of the backend network, this paper mainly improves the second and third additional layers. First, a 1*1 convolutional layer is used to increase the dimension, and then a 3*3 channel-by-channel convolution is used to replace the ordinary convolutional layer. For the second additional layer, the ECA attention mechanism is introduced after the channel-by-channel convolution. The activation function is H-Swish, and then a 1*1 point-by-point convolution is used for dimensionality reduction. The structure diagram of the improved additional convolutional layer is shown in the attached figure. Fig.13 shown.

[0097] Application of ECA attention mechanism and H-Swish activation function

[0098] The ECA attention mechanism and H-Swish activation function are introduced in the second additional layer to effectively capture the correlation between channels and improve the model's efficiency in utilizing features of different channels.

[0099] The benefit of using channel-by-channel convolution and point-by-point convolution instead of ordinary convolution is to reduce some parameters and floating-point calculations. The introduction of the ECA attention mechanism effectively captures the correlation between channels, improves the model's efficiency in utilizing different channel features, and helps optimize the model's performance. The H-Swish activation function is used because it is more friendly to high-dimensional channel expressions.

[0100] 4. Construction of lightweight pedestrian detection model MNetV3-Smallv2-SSD

[0101] Replacement and trimming of the backbone network: Replace the original backbone network VGG16, and trim the network layers after the 1*1 convolution layer after the 11th Bneck layer based on the MNetV3-Smallv2 network model, leaving only the hole convolution module, 11 Bneck layers and a 1*1 convolution layer as the backbone feature extraction network of the pedestrian detection model (see the attached network structure diagram) Figure 3 ), where the structure of the Bneck layer of the lightweight cross-layer skip connection module is shown in the attached figure. Figure 4 The MNetV3-Smallv2-SSD network structure is shown in the attached Fig.14 shown.

[0102] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0103] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A lightweight pedestrian detection method based on an improved SSD algorithm, characterized in that: include: Get the target input image; Using a lightweight pedestrian detection model to detect the target input image, and obtain a pedestrian recognition result; The lightweight pedestrian detection model is built with MobileNetV3-Small as the basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding additional convolutional layers after the basic convolutional network.

2. The lightweight pedestrian detection method based on the improved SSD algorithm according to claim 1 is characterized in that: The method for constructing the lightweight pedestrian detection model is as follows: S1. Backbone network construction: MobileNetV3-Small is selected as the basic network, and an improved MNetV3-Smallv2 network model is constructed by introducing SE and CA attention mechanisms; S2, image resizing: resize the input image from the original 224×224 to 300×300; S3. Application of dilated convolution module: Use the dilated convolution module to replace the first ordinary convolution layer in the basic network; the dilated convolution module is an expanded convolution layer with a dilation rate of (1, 2); S4, CA attention mechanism integration: Introduce the CoordinateAttention mechanism in the Bneck block; S5. Cross-layer skip connection design: According to the feature size and channel of the Bneck layer, a lightweight cross-layer skip module is designed, and the ECA attention mechanism is introduced; S6. Anchor frame customization optimization: The anchor frame is redesigned based on the Anchor mechanism, three aspect ratios are set, and the minimum prior frame size on the feature map is adjusted; S7, lightweight design of backend network: Improve the backend network of the SSD algorithm, use 1×1 convolution to increase the dimension, 3×3 channel-by-channel convolution for feature learning, and 1×1 point-by-point convolution for dimensionality reduction and information fusion; S8, ECA attention mechanism and H-Swish activation function application: Introduce ECA attention mechanism and H-Swish activation function in the additional layer; S9, lightweight pedestrian detection model construction: replace the original backbone network VGG16, cut the MNetV3-Smallv2 network, retain the key feature extraction module, and build a lightweight pedestrian detection model; S10. Experimental verification and performance evaluation: Experimental verification is performed on the Caltech pedestrian dataset to evaluate the model’s performance indicators such as mAP and FPS.

3. The lightweight pedestrian detection method based on the improved SSD algorithm according to claim 2 is characterized in that: The dilated convolution module in S3 is applied as follows: in the first 3×3 convolution layer in the Dilated Convolution block, the padding is 1, the convolution step is 2, and the dilation rate is 1; in the second 3×3 convolution layer, the padding is 2, the convolution step is 1, and the dilation rate is 2.

4. The lightweight pedestrian detection method based on the improved SSD algorithm according to claim 2 is characterized in that: The method of integrating the CA attention mechanism in S4 is: generating position weights by calculating the global average pooling of the horizontal and vertical coordinates of the input feature map, and then combining these weights with the original feature map to enhance the feature representation.

5. The lightweight pedestrian detection method based on the improved SSD algorithm according to claim 2 is characterized in that: The process of lightweight design of the back-end network in S7 is as follows: first use a 1×1 convolution layer in the additional layer to increase the dimension, then use a 3×3 channel-by-channel convolution to replace the ordinary convolution layer, and introduce the ECA attention mechanism and H-Swish activation function in the second additional layer, and finally use a 1×1 point-by-point convolution to reduce the dimension.

6. A lightweight pedestrian detection system based on an improved SSD algorithm, characterized in that: include: An image acquisition module, used to acquire a target input image; The image detection module is used to detect the target input image using a lightweight pedestrian detection model to obtain a pedestrian recognition result; the lightweight pedestrian detection model is constructed with MobileNetV3-Small as the basic network, combined with SE and CA attention mechanisms, and an improved SSD algorithm; the improved SSD algorithm further extracts feature information by adding an additional convolutional layer after the basic convolutional network.

7. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the lightweight pedestrian detection method based on the improved SSD algorithm according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the lightweight pedestrian detection method based on the improved SSD algorithm as described in any one of claims 1 to 5.