A method, apparatus, medium, and device for pedestrian target detection

Abstract

This invention discloses a method, apparatus, medium, and device for pedestrian target detection, relating to the field of image detection technology. This invention combines the fusion process of infrared and visible light images with the diffusion generation process of a diffusion model. Noise is added during the forward process to obtain infrared and visible light noisy images. During the reverse process, global semantic features and local edge texture features of both the infrared and visible light noisy images are continuously extracted at multiple levels, effectively preserving key feature information from both images. Based on this, denoising and fusion are performed to achieve high-quality fusion of infrared and visible light images for pedestrian target detection. This significantly improves the detection accuracy and robustness of pedestrian targets in complex scenes, providing more reliable technical support for applications such as intelligent monitoring and autonomous driving.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method, apparatus, medium and device for pedestrian target detection. Background Technology

[0002] Currently, with continuous social development, accelerated urbanization, and increased population density, traffic accidents, especially pedestrian injuries and deaths, have become a major social concern. The rapid increase in the number of motor vehicles has led to a corresponding increase in the frequency of traffic accidents. Pedestrians, as a vulnerable group in traffic accidents, have a high proportion of injuries and fatalities, especially at night or in low-light conditions, where they are more susceptible to injury due to a lack of adequate protection. Pedestrian detection technology, as one of the core technologies of autonomous driving systems, aims to accurately identify pedestrians on the road and take timely avoidance measures to ensure their safety. However, in complex environments such as low light or nighttime, the performance of pedestrian detection based on visible light images significantly decreases, especially when lighting conditions are insufficient or the field of view is limited, greatly reducing detection accuracy.

[0003] To address this challenge, infrared imaging, as an imaging technology with strong nighttime detection capabilities, is increasingly being applied to pedestrian detection tasks. Infrared images can provide clear target information in low light, nighttime, or inclement weather conditions. However, in some complex scenarios, the detail and texture information of infrared images are limited, and they are easily affected by background noise. Therefore, relying solely on infrared images for pedestrian detection is not always ideal. To solve this problem, the fusion technology of infrared and visible light images has gradually become a research hotspot in the field of pedestrian detection. By effectively fusing the two modalities, the advantages of visible light images in terms of detail and texture can be utilized to compensate for the shortcomings of infrared images in this regard, while fully leveraging the advantages of infrared images in low-light environments at night, thus improving the robustness and stability of the detection system under various environmental conditions. The fused image not only enhances the representation of the target but also suppresses the influence of background noise, improving the accuracy of pedestrian detection, especially under complex weather conditions.

[0004] In existing technologies, multispectral pedestrian detection methods improve detection accuracy by fusing infrared and visible light images. The core technology lies in effectively combining information from infrared and visible light images. Based on different fusion strategies, multispectral pedestrian detection methods are mainly divided into three categories: pixel-level fusion, feature-level fusion, and decision-level fusion. Pixel-level fusion directly fuses images at the image level, such as fusing low-frequency and high-frequency components separately after wavelet decomposition, or fusing at the image channel level. Feature-level fusion extracts image features from the two types of images using different neural network models and then fuses them. Decision-level fusion, based on the advantages of each type of image and considering the actual environment, performs decision-making fusion at the decision level with appropriate weights.

[0005] While existing multispectral pedestrian detection methods have improved detection accuracy through the fusion of infrared and visible light images, several technical challenges remain in practical applications. For instance, pixel-level fusion methods are prone to losing details in the fused image due to sensor noise or registration errors; feature-level fusion relies on manually designed feature extraction networks, failing to fully exploit the complementarity of cross-modal data; and decision-level fusion often requires prior knowledge for weight allocation, making it difficult to adapt to complex and changing scenarios. Therefore, existing multispectral pedestrian detection methods suffer from insufficient adaptability in dynamic environments and inadequate information fusion, resulting in poor detection accuracy. Summary of the Invention

[0006] Therefore, it is necessary to provide a pedestrian target detection method, apparatus, medium, and device to address the aforementioned technical problems.

[0007] The present invention adopts the following technical solution:

[0008] This invention provides a pedestrian target detection method. First, it acquires an infrared image and a visible light image to be detected. Then, it inputs these images into an improved diffusion model. Through the forward process of the improved diffusion model, noise is added to the infrared and visible light images respectively, resulting in an infrared noise image and a visible light noise image. Next, through the reverse process of the improved diffusion model, global semantic features and local edge texture features of the infrared noise image and the visible light noise image are extracted at multiple levels and fused layer by layer to obtain a denoised fused image. Finally, the denoised fused image is input into a pedestrian target detection model to obtain the pedestrian target detection result.

[0009] This invention provides a pedestrian target detection device, comprising:

[0010] The acquisition module is used to acquire the infrared and visible light images to be detected.

[0011] The forward noise addition module is used to input the infrared image and visible light image to be detected into the improved diffusion model. Through the forward process of the improved diffusion model, noise is added to the infrared image and visible light image to be detected respectively to obtain infrared noise image and visible light noise image.

[0012] The reverse denoising and fusion module is used to extract global image semantic features and local edge texture features of infrared noise images and visible light noise images through a multi-level reverse process of the improved diffusion model, and then fuse them layer by layer to obtain a denoised fused image.

[0013] The detection module is used to input the denoised and fused image into the pedestrian target detection model to obtain the pedestrian target detection result.

[0014] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described pedestrian target detection method.

[0015] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described pedestrian target detection method.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects:

[0017] This invention combines the fusion process of infrared and visible light images with the diffusion generation process of a diffusion model. During the forward diffusion process, noise is progressively added to obtain infrared and visible light noisy images. In the reverse generation process, global semantic features and local edge texture features of both the infrared and visible light noisy images are continuously extracted at multiple levels and fused layer by layer. This utilizes the denoising capabilities of the diffusion model and the extraction of global and local information at multiple levels for progressive fusion, effectively preserving the key feature information of both infrared and visible light images while achieving denoising fusion. This results in high-quality fusion of infrared and visible light images, fully leveraging the complementarity of cross-modal data for pedestrian target detection. This improves adaptability in dynamic environments and significantly enhances the detection accuracy and robustness of pedestrian targets in complex scenarios, providing more reliable technical support for applications such as intelligent monitoring and autonomous driving. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0019] Figure 1 This is a schematic diagram of a pedestrian target detection method provided by the present invention;

[0020] Figure 2 A schematic diagram of the overall process of DIVIF provided by the present invention;

[0021] Figure 3 A schematic diagram of a reverse fusion framework provided by the present invention;

[0022] Figure 4 A schematic diagram of a forward noise addition process provided by the present invention;

[0023] Figure 5 A schematic diagram of an IGCM module structure provided by the present invention;

[0024] Figure 6 This is a schematic diagram of a Lite Transformer framework provided by the present invention;

[0025] Figure 7 A schematic diagram of a YOLOv8 model network structure provided by the present invention;

[0026] Figure 8 A schematic diagram of an improved YOLOv8n framework for introducing MobileNetV4 into the backbone network provided by the present invention;

[0027] Figure 9 A schematic diagram of a C2f-GSC module structure provided by the present invention;

[0028] Figure 10 A schematic diagram of a YOLOv8n framework for introducing a C2f-GSC module into the neck network provided by the present invention;

[0029] Figure 11 A schematic diagram of a YOLOv8n framework for introducing an attention mechanism into the backbone network, provided by the present invention;

[0030] Figure 12 This invention provides a schematic diagram comparing image fusion using different models.

[0031] Figure 13 This is a schematic diagram of a pedestrian target detection device provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0033] Addressing the challenges of current multispectral pedestrian detection technologies, this invention, based on visible and infrared images, delves into pedestrian target detection methods, aiming to fully exploit the feature information of different spectral images to achieve efficient pedestrian detection in all-weather environments. This invention proposes an infrared-visible image fusion method based on a diffusion model. Through an iterative optimization mechanism of the diffusion process, it can more effectively preserve the texture details and thermal radiation features of multimodal images, while enhancing the robustness of feature representation, thereby solving the shortcomings of existing methods such as insufficient adaptability in dynamic environments and inadequate information fusion. To this end, this invention designs a pedestrian target detection model with superior performance, mainly including the following aspects:

[0034] (1) From the perspective of heterogeneous data feature fusion, and addressing the limitations of traditional fusion methods in cross-modal feature learning, this invention proposes an improved fusion algorithm based on a diffusion model. First, to solve the problem of multimodal image feature differences, a reversible global capture module is designed. By extracting and integrating global features from infrared and visible light images, it effectively balances detailed information with overall structural consistency, laying the foundation for subsequent fusion tasks. Second, to address the problem of insufficient local feature interaction, a lightweight local capture module is proposed. An attention mechanism is introduced into a lightweight network to achieve dynamic interaction and fusion of local features. By executing the two modules in parallel, features at different scales are fully utilized to achieve deep image fusion.

[0035] (2) For pedestrian detection tasks involving fused images, real-time performance is crucial, but existing algorithms generally suffer from high complexity and large memory consumption. While some lightweight networks improve detection speed by simplifying the structure, they often fail to maintain detection accuracy. To address this contradiction, this invention proposes an improved network based on YOLOv8n, achieving lightweight design while maintaining high efficiency. Specifically, firstly, lightweight modules are introduced into the YOLOv8n base model, effectively reducing network complexity; secondly, to balance accuracy and speed, an attention mechanism module is embedded in the network, significantly improving the model's pedestrian detection performance in low-light environments by enhancing key feature extraction capabilities and suppressing interference from non-critical information.

[0036] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Figure 1 This is a schematic diagram of a pedestrian target detection method according to the present invention, which specifically includes the following steps:

[0038] S101: Acquire the infrared and visible light images to be detected;

[0039] S102: Input the infrared image and visible light image to be detected into the improved diffusion model. Through the forward process of the improved diffusion model, noise is added to the infrared image and visible light image to be detected respectively to obtain the infrared noise image and the visible light noise image.

[0040] S103: The global image semantic features and local edge texture features of infrared noise images and visible light noise images are extracted in multiple levels through the reverse process of the improved diffusion model and then fused layer by layer to obtain a denoised fused image.

[0041] S104: Input the denoised and fused image into the pedestrian target detection model to obtain the pedestrian target detection result.

[0042] The server mentioned in this invention can be a server set up on a business platform, or a device such as a desktop computer or laptop computer capable of executing the solution of this invention. For ease of explanation, the following description will only focus on the server as the executing entity.

[0043] Visible light is electromagnetic wave perceptible to the human eye, with a wavelength range between 400 and 760 nanometers. Visible light sensors can capture electromagnetic waves within this wavelength range and convert the received light energy into electrical signals, enabling electronic devices to recognize light information. The resulting images accurately reflect the color characteristics of the target object. Visible light images possess high spatial resolution, high contrast, and rich color information, clearly presenting the texture, edge contours, and color details of the target, highly replicating the human visual experience. However, their imaging quality is heavily dependent on lighting conditions: under ideal illumination, visible light images can capture complete visual details of a scene; but at night or in low-light environments, their ability to acquire target information decreases significantly, especially in low-visibility conditions such as nighttime or heavy fog, where they struggle to penetrate opaque objects, leading to reduced image quality and missing target information. This limitation restricts their application in low-light scenarios.

[0044] The imaging principles of infrared and visible light images differ significantly. Infrared radiation refers to electromagnetic waves with wavelengths ranging from 0.78 micrometers to 1000 micrometers. Infrared images are generated by mapping the intensity of infrared radiation emitted by a target object into a grayscale distribution using an infrared sensor. Its imaging mechanism is based on the surface temperature of the object; the higher the temperature, the stronger the infrared radiation. Compared to visible light images, infrared imaging has stronger environmental adaptability, is less affected by lighting conditions, and can effectively detect targets even in obstructed or nighttime conditions, enabling all-weather operation. However, infrared images have relatively low spatial resolution, and texture details are not as clear, which may lead to a blurry effect during observation and does not conform to human visual perception habits. These characteristics give infrared imaging unique advantages in specific scenarios, but also present certain limitations.

[0045] Therefore, fusing infrared and visible light images can fully leverage the complementary advantages of the two modalities: infrared images provide reliable thermal radiation information, ensuring target detection capabilities under low light or occlusion conditions; visible light images supplement color and detail information, enhancing target discernibility. This fusion strategy can significantly improve the accuracy of target detection, especially in complex scenes (such as nighttime, fog, etc.).

[0046] Therefore, the server of the business platform can first obtain pre-collected historical infrared images and historical visible light images as sample images to train the model used for image fusion and apply it to pedestrian target detection.

[0047] In one or more embodiments of the present invention, an infrared image and visible light image fusion algorithm based on a diffusion model is proposed to address the shortcomings of single-modal images in information representation.

[0048] Based on the diffusion model framework, a reverse fusion network structure is proposed in the reverse process. Specifically, the reverse fusion network structure includes an Invertible Global Capture Module (IGCM) and a Lightweight Local Capture Module (LLCM). The IGCM achieves cross-modal feature decoupling and reconstruction through reversible transformation, which can separate multimodal image noise and retain common global image semantic features. The LLCM can extract local edge texture features of multimodal images and dynamically inject them during the fusion process, ultimately reconstructing high-quality features after fusion to obtain a high-quality denoised fused image. This design gradually reduces modal differences through iterative optimization, surpassing the unidirectional denoising process of the standard diffusion model.

[0049] The Denoising Diffusion Probabilistic Model (DDPM) excels in generative tasks due to its superior generative capabilities. Its core principle is to gradually add noise to the prior data distribution using forward diffusion, transforming it into random noise, and then using multi-step optimization in backward diffusion to gradually recover the original data, ensuring the generated results are realistic and rich in detail. In one or more embodiments of this invention, the diffusion model used as the basis for improvement can be the denoising diffusion probabilistic model.

[0050] This invention proposes an infrared-visible image fusion algorithm (DIVIF) based on an improved diffusion model. DIVIF employs a multi-stage alternating optimization strategy, which focuses on enhancing the details and features of both infrared and visible images during the iteration process to form a unified feature representation and gradually generate a high-quality fused image.

[0051] Figure 2 This is a schematic diagram of the overall process of DIVIF in this invention. DIVIF includes a forward noise diffusion process and a reverse continuous fusion process. In the forward diffusion process, Gaussian noise is gradually added to the infrared and visible light images until it is completely destroyed into pure noise. In the reverse continuous fusion process, with the assistance of the infrared and visible light images, a fusion operation of heterogeneous images is performed at each time step to gradually generate a fused image from pure noise. The specific auxiliary process is as follows: Taking time step t→t-1 as an example, the infrared noise image and the visible light noise image generated by the forward process are first received. The global image semantic features of each modality are reconstructed through IBL-Block IGCM, while complementary local edge texture features are extracted by LLCM.

[0052] In the reverse continuous fusion process of DIVIF, IBL-Block (Integrated Base and LocalBlock) works collaboratively with the IGCM and LLCM modules to achieve multi-level feature fusion of infrared and visible light images. IGCM is responsible for extracting global contextual information (such as scene structure and illumination distribution) from both modalities, while LLCM focuses on capturing local detail features (such as edge and texture differences). Through the dynamic feature integration mechanism of IBL-Block, they generate optimized fused reconstructed images step-by-step. By iteratively optimizing the fusion path, the temporal consistency and feature integrity of the reconstruction results are gradually improved. The reverse fusion process is as follows: Figure 3 As shown, Figure 3 This is a schematic diagram of a reverse fusion framework in this invention.

[0053] Depend on Figure 3 As can be seen, in one or more embodiments of the present invention, the reverse part of the improved diffusion model may include a denoising feature extraction layer and a feature fusion layer; the denoising feature extraction layer includes a parallel first branch and a second branch, both of which include multiple serially connected comprehensive extraction modules, the first branch being used to receive visible light noise images, and the second branch being used to receive infrared noise images;

[0054] Each integrated extraction module may include parallel reversible global capture modules and lightweight local capture modules, and the feature fusion layer may include multiple cascaded fusion layer modules. In the fusion layer, infrared features provide target saliency guidance, while visible light features provide texture detail constraints; the final fused image retains both modal advantages.

[0055] Furthermore, the reversible global capture module includes a reversible neural network model, and the lightweight local capture module may include a Lite Transformer model.

[0056] The IGCM module proposed in this invention focuses on extracting global features from infrared and visible light images, capturing large-scale structural information and enhancing the model's ability to perceive the consistency of global features, providing stable global support for subsequent feature refinement operations. Building upon this, the LLCM module utilizes an attention mechanism to dynamically interact and fuse local features from infrared and visible light images, capturing fine-grained feature correlations and significantly improving the detail fidelity of the fused image.

[0057] Throughout the backdiffusion process, the IGCM and LLCM modules execute in parallel, forming a continuous optimization process that gradually achieves in-depth refinement and fusion of image features. Within each IBL-Block, the IGCM and LLCM adopt a parallel-serial hybrid structure. After cascading processing of multiple IBL-Blocks, multi-level global image semantic feature extraction and local edge texture feature extraction are achieved for the bimodal noisy image. The bimodal global image semantic feature extraction and local edge texture feature extraction are then weighted and fused in the Fusion Layer.

[0058] In one or more embodiments of the present invention, for the fusion process of the feature fusion layer, a Fusion layer module connected to the last comprehensive extraction module in the first branch and the second branch is used as the first Fusion layer module. The first Fusion layer module is used to receive the global image semantic features and local edge texture features of the infrared noise image and the global image semantic features and local edge texture features of the visible light noise image obtained by the last comprehensive extraction module in the first branch and the second branch, and perform feature fusion to obtain an initial denoised fused image, which is then transmitted to the second Fusion layer module. The second Fusion layer module performs feature fusion based on the initial denoised fused image and the features obtained by the second-to-last comprehensive extraction module in the first and second branches. Each Fusion layer module and the comprehensive extraction modules in the first and second branches sequentially form a reverse-order denoised fusion structure. (See reference here.) Figure 3 The indication, Figure 3The example shown uses three comprehensive extraction modules in each branch and three fusion layer modules in the feature fusion layer.

[0059] Through this alternating optimization mechanism, DIVIF achieves modal complementarity between infrared and visible light images within a unified diffusion framework. The resulting fused image possesses rich detail and high-quality spectral information, thus significantly improving the performance of the fused image.

[0060] Given a pair of visible light and infrared images, the goal of the DIVIF model is to generate a high-quality fused image of a target by progressively extracting and fusing multimodal features. To achieve this goal, the DIVIF model employs the core idea of ​​a diffusion model, which involves progressively disrupting the data distribution of the images through a forward process and reconstructing the fused image during the reverse generation process.

[0061] During the forward process, Gaussian noise is gradually added to the visible light and infrared images, causing their data distribution to gradually degrade. After T time steps... It eventually degenerates into a near-pure Gaussian distribution. This process can be represented as: .

[0062] in, , Indicates the first Infrared and visible light noise images of the steps, Indicates the first Step images, It is the time step The noise attenuation coefficient, It is the identity matrix. Variance By controlling the intensity of the added noise, as the time step increases, the noise gradually becomes dominant, eventually leading to... It is close to a pure Gaussian distribution.

[0063] The design of the forward process has the following important implications: First, by progressively adding noise, the improved diffusion model can transform the complex original image distribution into a simple Gaussian distribution, providing a clear objective for the reverse process. Second, the reversibility of the forward process provides theoretical support for the reverse process, enabling the model to reconstruct a high-quality fused image through progressive denoising. Finally, the noise addition strategy in the forward process can effectively balance feature preservation and distribution degradation, avoiding premature loss of key information. In the DIVIF model, the core of the forward process lies in the noise attenuation coefficient. The design is reasonable. This process ensures the smoothness and controllability of noise addition, thereby preserving key features of multimodal images while preventing image distribution degradation. Subsequent experiments validated the effectiveness of the forward process, providing a solid foundation for subsequent feature extraction and fusion. The final fused image exhibits excellent performance in both visual quality and information integrity. Figure 4 This is a schematic diagram of a forward noise addition process in this invention.

[0064] For the reverse process, based on the noisy image generated in the forward process, the improved diffusion model progressively removes noise and dynamically fuses the multimodal features of the infrared and visible light images through a reverse fusion network. This reverse process takes the noisy image as input and, through a reversible local feature extraction module and a global lightweight extraction module, achieves hierarchical interaction of cross-modal features while denoising and reconstructing, ultimately generating a high-quality fused image that combines thermal radiation information and texture details. Through an end-to-end training strategy, the reverse fusion network can effectively learn the complementary relationships between multimodal features.

[0065] Considering the importance of infrared and visible light images in the fusion of detailed features (such as edges and textures), the reverse process must minimize information loss while ensuring feature complementarity and detail fidelity between modalities. In one or more embodiments of the present invention, the reversible global capture module IGCM is based on the architecture of an invertible neural network (INN). The unique reversible network architecture allows for the establishment of a bidirectional mapping relationship between input and output features, thereby ensuring lossless information transmission. Figure 5 This is a schematic diagram of an IGCM module structure according to the present invention. Figure 5 As can be seen, the IGCM module has a bidirectional mapping relationship between input and output features. In practical applications, only its forward process can be used for denoising and fusion. Figure 5 "in " indicates corresponding addition", " indicates corresponding multiplication.

[0066] At diffusion time step t, the IGCM module receives infrared image features from the previous time step. or visible light image features As input, global semantic features are extracted, specifically through feature fusion extraction using multiple reversible blocks. Each reversible block may include three nonlinear mapping functions. , and .

[0067] in, , and These are learnable neural networks, used for feature extraction and fusion, respectively. For and The bottleneck residual block (BRB) in the MobileNetV2 module can be used as their network architecture. Through this process, IGCM can effectively extract global image semantic information from different modalities, laying the foundation for subsequent reconstruction processes.

[0068] For the lightweight local capture module (LLCM), in one or more embodiments of the present invention, the LiteTransformer model may be used. Figure 6 This is a schematic diagram of a Lite Transformer framework in this invention. Figure 6 The “CSATransformer Block” indicates that a Convolutional Self-Attention (CAS) mechanism is introduced in the first stage of the network, while the “RASA Transformer Block” indicates that a Recursive Atrous Self-Attention (RASA) mechanism is introduced in a subsequent stage of the network.

[0069] The Lite Transformer architecture employed in this invention is an existing and mature lightweight visual Transformer model. Its core design concept lies in integrating the powerful representation capabilities of the Transformer with the efficient computational characteristics of the multilayer perceptron. This model achieves a balance between performance and efficiency through three key technological innovations: First, it employs a hierarchical Transformer encoder structure, simultaneously capturing local details and global contextual information of the image through a multi-scale feature extraction mechanism; second, it abandons traditional positional encoding schemes, effectively solving the performance degradation problem caused by resolution inconsistencies between training and testing phases; finally, through a simplified MLP decoder design, it achieves intelligent aggregation of cross-layer features, organically combining local receptive fields with global attention mechanisms. This unique architectural design not only significantly reduces computational complexity but also maintains excellent visual feature representation capabilities, providing an efficient solution for various visual tasks.

[0070] Based on the improved diffusion model described above, the server can input the previously acquired historical infrared images and historical visible light images into the improved diffusion model to obtain a fused image and determine a first loss to train the diffusion model, thus obtaining an image fusion model.

[0071] Regarding the loss in image fusion, DIVIF aims to generate a fused image through an end-to-end learning process and a series of fusion steps. Therefore, traditional content losses, including intensity loss and texture loss, are considered in the design of the loss function.

[0072] Intensity loss is used to constrain the global brightness consistency of the fusion result to ensure the accuracy of the reconstructed image in terms of overall intensity; while texture loss is used to preserve and enhance local texture details in the image to ensure the richness of texture detail in the fused image. Content loss is shown in the following equation: .

[0073] In the formula, For content loss during image fusion, The intensity loss during the image fusion process, Texture loss during image fusion process, This is a loss balancing parameter used to maintain the balance between the intensity loss and texture loss.

[0074] Strength loss and texture loss The definition is as follows:

[0075] , .

[0076] In the formula, Image height, Image width, To merge images, To input an infrared image, To input a visible light image, The Sobel gradient operator is used to measure fine-grained texture information of an image. To find the maximum function, for Norm, Calculated for absolute values.

[0077] After obtaining the fused image as described above, pedestrian target detection can be further performed based on the fused image. In one or more embodiments of the present invention, the fused image can be input into a YOLOv8 model as the detection model to obtain pedestrian target prediction results and determine a second loss to train the detection model, thus obtaining a pedestrian detection model. YOLOv8 provides five different model sizes: n, s, m, l, and x. The accuracy of the model increases with the increase of the model size. In order to balance accuracy and computational resource consumption in practical applications, a smaller model is usually selected as the benchmark based on the computing power of the deployed network and the application scenario. This is especially true for the n and s models. Figure 7 This is a schematic diagram of a YOLOv8 model network structure according to the present invention, which mainly includes four parts: input layer, backbone network, neck network, and head. The modules in the four parts are all existing modules and will not be described in detail here.

[0078] Furthermore, in one or more embodiments of the present invention, considering the dual requirements of real-time performance and accuracy for pedestrian detection tasks in multimodal fusion images, the backbone and neck networks of the YOLOv8 model can be lightweighted and improved by introducing attention mechanisms and loss function improvements, focusing on optimizing multimodal fusion feature extraction and pedestrian target localization capabilities. This ensures pedestrian detection accuracy while reducing network computational complexity and improving detection efficiency to meet the performance requirements of embedded devices in practical application scenarios.

[0079] The Input layer is primarily responsible for image preprocessing and data augmentation, providing high-quality input for subsequent feature extraction. The Backbone is the core module in the YOLOv8 model used for feature extraction; its main function is to extract multi-level feature information from the input image for use by subsequent networks. The Neck is the key part connecting the Backbone and Head in the YOLOv8 model, mainly responsible for further fusing and optimizing the features extracted by the Backbone to improve the learning ability of the Head layer in specific tasks. The Head layer is responsible for the final object classification and localization tasks.

[0080] While the YOLOv8 model has demonstrated outstanding performance in object detection, its large number of parameters limits its potential for optimization in terms of computational efficiency and resource consumption. Especially when processing fused images, the large input size means that directly deploying them on mobile devices leads to a significant performance degradation, making it difficult to meet the demands of real-time detection tasks. To address these issues, this invention focuses on lightweighting the backbone and neck networks of the YOLOv8n model, aiming to reduce model complexity and improve computational efficiency. Furthermore, by introducing an attention mechanism, it reduces the computational burden on the network while simultaneously improving detection accuracy.

[0081] Based on this, in one or more embodiments of the present invention, a pedestrian target detection model based on the YOLOv8 model can be obtained by replacing the combination of each convolution and C2f module in the backbone network of the YOLOv8n model with a MobileNetV4 module. Figure 8 This is a schematic diagram of an improved YOLOv8n framework that incorporates MobileNetV4 into the backbone network of this invention.

[0082] MobileNetV4 is a lightweight network designed to optimize model performance on mobile and embedded devices while maintaining high efficiency. It excels in image classification tasks. MobileNetV4 employs efficient depthwise separable convolutions and channel attention mechanisms, significantly reducing the number of parameters while maintaining high feature extraction capabilities in complex scenes. Furthermore, MobileNetV4 exhibits faster inference speeds on mobile devices, making it particularly suitable for computationally demanding tasks such as real-time pedestrian detection. Therefore, this invention selects MobileNetV4 as the backbone network of YOLOv8n to achieve a lightweight model design and improve computational efficiency while ensuring that detection accuracy remains unaffected.

[0083] MobileNetV4 has three main versions: MobileNetV4-Small, MobileNetV4-Medium, and MobileNetV3-Large, which correspond to versions with low to high computational and storage requirements, respectively. This invention aims to make the algorithm as lightweight as possible, so MobileNetV4-Small was chosen. In the UIB module, the ExtraDW module was selected because, in pedestrian detection tasks, the ExtraDW module can help the network maintain high performance in low-resolution and multi-scale situations, while reducing computational resource consumption and improving real-time detection results.

[0084] In processing the aforementioned fused image, the input fused image is first processed through a backbone network for feature extraction. The main function of this backbone network is to convert the spatial information of the image into channel information, laying the foundation for subsequent feature processing.

[0085] However, during the process of progressively converting spatial information into channel information, the backbone may experience some loss of semantic information. To address this issue, in one or more embodiments of the present invention, multiple Bottleneck modules in the C2f module of the YOLOv8n network's neck network can be replaced with multiple GS Bottleneck modules, resulting in a C2f-GSC module introduced into the neck network. This GS Bottleneck module is derived from the GSConv Bottleneck module. Figure 9 This is a schematic diagram of a C2f-GSC module structure according to the present invention. Figure 10 This is a schematic diagram of a YOLOv8n framework that introduces a C2f-GSC module into the neck network in this invention.

[0086] GSConv preserves the connections between channels. In the Neck part of the network, the feature map after Backbone processing reaches its maximum number of channels, while its width and height dimensions are minimized. Introducing GSConv at this point effectively reduces information redundancy, lowers the computational burden, and accelerates the inference process. The time complexity of GSConv is shown in the following equation: .

[0087] In the formula, These represent the width and height of the output feature map, respectively. These are the number of channels in the input feature map and the output feature map, respectively.

[0088] The GSConv module achieves efficient feature extraction and cross-channel information interaction by combining ordinary convolution (Conv), depthwise separable convolution (DWConv), and channel shuffle operations. First, the input image is processed through ordinary convolution to generate an initial feature map. This initial feature map is then processed through depthwise separable convolution to obtain another set of feature maps. Next, a specific fusion strategy combines the feature maps extracted by ordinary convolution and depthwise separable convolution to fully utilize the information from both. Finally, a channel shuffle operation is performed on the fused feature map, enabling cross-channel interaction between the feature information extracted by ordinary convolution and that extracted by depthwise separable convolution.

[0089] This design not only maintains detection accuracy but also significantly reduces computational costs. GSConv achieves effective fusion of cross-channel information through channel shuffling operations, thereby improving the comprehensiveness of feature extraction while reducing computational load. The computational cost of GSConv is shown in the following formula:

[0090] .

[0091] In the formula, The kernel size is [size]. Indicates the number of groups.

[0092] Furthermore, attention mechanisms enable neural networks to automatically learn and selectively focus on key components of the input data. Attention mechanisms can be categorized by the dimension of focus, such as spatial, channel, temporal, and mixed domains. In object detection tasks, attention mechanisms help models selectively focus on regions or features of interest, allocate appropriate resources to each convolutional channel, and suppress irrelevant information. In pedestrian detection using fused infrared and visible light images, attention mechanisms effectively suppress background noise and enhance focus on pedestrian features, thereby improving detection accuracy. Therefore, by introducing attention mechanisms, noise from different modalities can be effectively filtered and corrected before feature fusion, allowing the model to better cope with complex environmental changes and improve the final detection performance.

[0093] Currently, there are several mature attention mechanisms, such as the SE (Squeeze-and-Excitation) attention mechanism, the CBAM (Convolutional Block Attention Module) attention mechanism, and the EMA (Exponential Moving Average) attention mechanism.

[0094] In one or more embodiments of the present invention, a lightweight backbone network design is achieved by selecting MobileNetV4 based on the YOLOv8n network. Furthermore, a lightweight convolutional module GSConv is introduced into the neck network to replace the traditional convolution in the C2f stage of the neck network, thus achieving a lightweight neck network design. However, since the improved lightweight network often loses some detection accuracy, one of the aforementioned CA attention module, CBAM attention module, and EMA attention module can be selectively introduced into the end of the backbone network to enhance its feature learning capabilities, or other attention mechanisms can be introduced. Figure 11 This is a schematic diagram of the YOLOv8n framework for introducing an attention mechanism into the backbone network in this invention.

[0095] Based on the detection model based on the YOLOv8 model obtained in one or more of the above embodiments, the server can input the fused image into the detection model to obtain the pedestrian target prediction result and determine the second loss to train the detection model, thereby obtaining the pedestrian detection model.

[0096] For the second loss, YOLOv8 continues the CIoU loss calculation method of YOLOv5 in the design of the loss function, as shown in the following formula:

[0097] .

[0098] In the formula, CIoU loss for the ground truth bounding box and the predicted bounding box. The IOU loss is calculated for the ground truth bounding box and the predicted bounding box. for b and The Euclidean distance between them To predict the center coordinates of the bounding box, The center coordinates of the true bounding box. Let be the diagonal length of the minimum bounding rectangle between the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. A parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box.

[0099] Because YOLOv8 introduces an anchor-free, center-point-based approach, the model's output changes from predicting the bounding box coordinates to predicting the offsets of the left, top, right, and bottom borders of the target box from the target's center point. To synchronize with the anchor-free mechanism and further improve the model's generalization ability, YOLOv8 adds the Distribution Focal Loss (DFL) loss.

[0100] However, its accuracy and stability may be insufficient when dealing with target occlusion or blurred boundaries. CIoU mainly considers overlapping regions, center point distance, and aspect ratio when calculating bounding box regression. In addition, although DFL can accelerate the model's focusing on the target location, it may not be able to fully optimize the localization accuracy of the bounding box in complex scenes.

[0101] To address these issues, one or more embodiments of this invention propose employing an improved MPDIoU (Minimum Point Distance IoU) bounding box regression loss function. MPDIoU, by comprehensively considering the overlapping areas of the bounding boxes, the distance between their center points, and the minimum point distance, can more accurately measure the difference between the predicted and ground truth bounding boxes.

[0102] Specifically, the server can input the fused image into a detection model based on the YOLOv8 model to obtain the predicted bounding box corresponding to the pedestrian target. Based on the preset ground truth bounding box and the predicted bounding box of the pedestrian, the second loss is determined by the following formula:

[0103] ,

[0104] ;

[0105] in, This is the second loss. The coordinate frame of the true bounding box, ( , ) represents the coordinates of the top-left corner of the true bounding box. , () represents the coordinates of the bottom right corner of the actual bounding box. To predict the coordinates of the bounding box, ( , ) represents the coordinates of the top-left corner of the predicted bounding box. , () represents the coordinates of the bottom right corner of the predicted bounding box. For the true bounding box, To predict the bounding box, The minimum Euclidean distance between the pairs of diagonal vertices of the true bounding box and the predicted bounding box. The maximum Euclidean distance between the pairs of diagonal vertices of the true bounding box and the predicted bounding box. The coordinates of the top-left corner of the actual bounding box. The coordinates of the bottom right corner of the true bounding box. To predict the coordinates of the top left corner of the bounding box, To predict the coordinates of the top left corner of the bounding box.

[0106] After obtaining the trained image fusion model and pedestrian detection model, in practical applications, the server can acquire the infrared image and visible light image to be detected and input them into the image fusion model. The infrared image and visible light image are noise-added separately through the forward process of the diffusion model to obtain infrared noise image and visible light noise image. The global image semantic features and local edge texture features of the infrared noise image and the global image semantic features and local edge texture features of the visible light noise image are extracted in multiple levels through the reverse process and fused layer by layer to obtain a denoised fused image. The pedestrian detection model is then used to detect pedestrian targets on the denoised fused image to obtain the pedestrian target detection result.

[0107] based on Figure 1 The pedestrian target detection method shown in this invention combines the fusion process of infrared and visible light images with the diffusion generation process of a diffusion model. During the forward diffusion process, noise is gradually added to obtain infrared and visible light noisy images. In the reverse generation process, global semantic features and local edge texture features of both the infrared and visible light noisy images are continuously extracted at multiple levels and fused layer by layer. This method utilizes the denoising capabilities of the diffusion model and the extraction of global and local information at multiple levels for gradual fusion, effectively preserving the key feature information of both infrared and visible light images while achieving denoising fusion. It realizes high-quality fusion of infrared and visible light images, fully exploiting the complementarity of cross-modal data for pedestrian target detection. This improves adaptability in dynamic environments and significantly enhances the detection accuracy and robustness of pedestrian targets in complex scenes, providing more reliable technical support for applications such as intelligent monitoring and autonomous driving.

[0108] To address the shortcomings of traditional fusion methods in cross-modal feature learning, an improved diffusion model-based fusion algorithm, DIVIF, is proposed. This method fully leverages the powerful generative capabilities of the diffusion model to improve the clarity and information completeness of the fused image. In the reverse process, DIVIF utilizes a reversible global capture module to extract features from infrared and visible light images, thereby acquiring large-scale structural information and enhancing global consistency. Secondly, a lightweight local capture module dynamically interacts and fuses local features through an attention mechanism, improving detail fidelity and spatial consistency. By executing both in parallel, deep refinement and fusion of image features are achieved.

[0109] Then, addressing the issues of large parameter count, insufficient detection accuracy, and high requirements for model accuracy and real-time performance in low-light traffic scenarios, a pedestrian target detection network based on YOLOv8n is proposed. This model can be lightweighted by using MobileNetV4 and GSConv in the backbone and neck network parts respectively. Furthermore, an EMA attention mechanism can be introduced, and the network loss function can be replaced with MPDIoU, maximizing detection accuracy while maintaining model lightweightness.

[0110] The present invention also provides embodiments for applying the present invention to verify the superior performance of the proposed reversible global capture module (IGCM) and lightweight local capture module (LLCM) in feature extraction, as well as the superior performance of the pedestrian target detection model.

[0111] For the image fusion part, this embodiment selected several representative multispectral detection networks for testing on the MSRS and LLVIP datasets. The experimental results are shown in Tables 1 and 2. The algorithms compared include DenseFuse, FusionGAN, RFN-Nest, SwinFusion, and DATFuse. Six widely used evaluation metrics—Standard Deviation (SD), Visual Information Fidelity (VIF), Average Gradient (AG), Information Entropy (EN), Gradient-based FusionPerformance (Qabf), and Spatial Frequency (SF)—were used to evaluate the quality of different methods on the datasets. Higher values ​​for these metrics indicate better quality fused images.

[0112] Table 1. Quantitative comparison on the MSRS dataset

[0113]

[0114] Table 2 Quantitative comparison on the LLVIP dataset

[0115]

[0116] The data shows that the DIVIF proposed in this invention performs best in almost all metrics, significantly outperforming other models, indicating that it has significant advantages in image detail preservation and sharpness.

[0117] To qualitatively analyze the visual effects of different fusion methods, this embodiment also conducted a visual comparison experiment. Typical images of roads under different daytime and nighttime conditions were selected from the dataset as test samples, and the experimental results were analyzed in detail. Figure 12 This is a schematic diagram comparing different model image fusion methods in this invention. Figure 12 The first two rows correspond to daytime, and the last two rows correspond to nighttime. The first two images in each pair are the original visible light image and the infrared image, and the last image corresponds to the method of this invention. The rest correspond to the fusion results of DenseFuse, FusionGAN, RFN-Nest, SwinFusion, and SwinFusion, respectively.

[0118] The comparison results show that the fusion algorithm proposed in this invention performs particularly well in dark environments. In visible light images, since extremely dark scenes are not present, various fusion methods can generate fused images with relatively ideal visual effects. However, in complex areas with local darkness, traditional fusion methods often struggle to clearly preserve the texture edge information of objects, resulting in significant loss of detail. Thanks to the improved diffusion model proposed in this invention, the fused image can fully retain the rich information and features of the source image, allowing the texture details in dark areas to be clearly presented, while pedestrian targets hidden in the darkness can also be accurately identified. Furthermore, the fused image generated by this method has a lower noise level, further improving the overall image quality. Experimental results demonstrate that the method of this invention exhibits stronger adaptability and generalization ability in complex dynamic lighting environments, especially in processing local dark areas, significantly improving image detail and target recognition, providing reliable technical support for image fusion tasks in low-light environments.

[0119] For the final pedestrian target detection, this embodiment made several improvements to the YOLOv8n network and compared it with other series of single-stage object detection networks YOLO and the two-stage object detection network Faster R-CNN. The experimental results are shown in Table 3, where "YOLOv8n-MGEL" represents the pedestrian target detection model after making the above-mentioned improvements to the YOLOv8n network.

[0120] Table 3 Comparison of results from different target detection networks

[0121]

[0122] Through in-depth analysis of the results, Faster R-CNN, as a classic two-stage detection algorithm, performs well in terms of accuracy and precision. However, due to its more complex model structure, its parameter count is approximately 25 times that of the method proposed in this paper. Single-stage detection algorithms such as YOLOv5n and YOLOv7n each have their advantages, but all indicators are lower than YOLOv8n. YOLOv8n is lower than the algorithm proposed in this paper in terms of both accuracy and model size. In contrast, the improved algorithm proposed in this paper achieves an mAP of 95.3%, which is 1.7% higher than the mAP of the basic YOLOv8n. In summary, the improved YOLOv8n-MGEL fusion image pedestrian detection algorithm of this invention maintains high detection accuracy while compressing the network parameter size, achieving a more lightweight model design, and fully demonstrating its advantages in pedestrian detection applications based on infrared and visible light fusion images.

[0123] The above describes a pedestrian target detection method provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding pedestrian target detection device, such as... Figure 13 As shown.

[0124] Figure 13 A schematic diagram of a pedestrian target detection device provided by the present invention includes:

[0125] The acquisition module 201 is used to acquire the infrared image and the visible light image to be detected;

[0126] The forward noise addition module 202 is used to input the infrared image and visible light image to be detected into the improved diffusion model, and add noise to the infrared image and visible light image to be detected respectively through the forward process of the improved diffusion model to obtain infrared noise image and visible light noise image;

[0127] The reverse denoising fusion module 203 is used to extract global image semantic features and local edge texture features of infrared noise images and visible light noise images through a multi-level reverse process of the improved diffusion model, and fuse them layer by layer to obtain a denoised fused image.

[0128] The detection module 204 is used to input the denoised and fused image into the pedestrian target detection model to obtain the pedestrian target detection result.

[0129] Specific limitations regarding the pedestrian target detection device can be found in the limitations of the pedestrian target detection method described above, and will not be repeated here. Each module in the aforementioned pedestrian target detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0130] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided pedestrian target detection method.

[0131] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The provided pedestrian target detection method.

[0132] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0133] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A pedestrian object detection method, characterized by, The method comprises the following steps: acquiring an infrared image and a visible light image to be detected; inputting the infrared image and the visible light image to be detected into an improved diffusion model, adding noise in the infrared image and the visible light image to be detected respectively through a forward process of the improved diffusion model to obtain an infrared noise image and a visible light noise image; extracting global image semantic features and local edge texture features of the infrared noise image and global image semantic features and local edge texture features of the visible light noise image through a backward process of the improved diffusion model in multiple levels and fusing them layer by layer to obtain a denoising fusion image; inputting the denoising fusion image into a pedestrian target detection model to obtain a pedestrian target detection result; the backward part of the improved diffusion model comprises a denoising feature extraction layer and a feature fusion layer; the denoising feature extraction layer comprises a first branch and a second branch in parallel, and the first branch and the second branch each comprise a plurality of serially connected comprehensive extraction modules, the first branch is used for extracting global image semantic features and local edge texture features of the visible light noise image through a plurality of serially connected comprehensive extraction modules in multiple levels, and the second branch is used for extracting global image semantic features and local edge texture features of the infrared noise image through a plurality of serially connected comprehensive extraction modules in multiple levels; each comprehensive extraction module comprises a reversible global capture module and a lightweight local capture module in parallel; the reversible global capture module comprises a reversible neural network model, and the lightweight local capture module comprises a Lite Transformer model; the feature fusion layer comprises a plurality of serially connected Fusion layer modules; a Fusion layer module connected with the last comprehensive extraction module in the first branch and the second branch is taken as a first Fusion layer module; the first Fusion layer module is used for receiving global image semantic features and local edge texture features of the infrared noise image and global image semantic features and local edge texture features of the visible light noise image obtained by the last comprehensive extraction module in the first branch and the second branch, fusing the features to obtain an initial denoising fusion image, and then transmitting the initial denoising fusion image to a next Fusion layer module, so that the next Fusion layer module fuses features according to the initial denoising fusion image and features obtained by a penultimate comprehensive extraction module in the first branch and the second branch, and the Fusion layer modules and the comprehensive extraction modules in the first branch and the second branch form a denoising fusion structure in reverse order in sequence.

2. The pedestrian object detection method of claim 1, wherein, training the improved diffusion model, specifically comprising: inputting historical infrared images and historical visible light images into the improved diffusion model, adding noise in the historical infrared images and the historical visible light images respectively through a forward process to obtain historical infrared noise images and historical visible light noise images; extracting global image semantic features and local edge texture features of the historical infrared noise images and global image semantic features and local edge texture features of the historical visible light noise images through a backward process in multiple levels and fusing them layer by layer to obtain historical denoising fusion images; The content loss of the image fusion process is determined according to the deviation between the fusion image and the source image based on the historical denoising fusion image by the following formula as a first loss, and the improved diffusion model is trained with the optimization target of minimizing the first loss: ； ， ； wherein, is a content loss for the image fusion process, is an intensity loss for the image fusion process, is a texture loss for the image fusion process, is a loss balancing parameter, is an image height, is an image width, is a historical denoised fused image, is a historical infrared image, is a historical visible light image, is a Sobel gradient operator for measuring fine-grained texture information of an image, is a max function, is is a norm, denotes taking an absolute value.

3. The pedestrian object detection method of claim 1, wherein, The pedestrian target detection model is obtained by replacing the combination of each convolution and C2f module in the backbone network of the YOLOv8n network with a MobileNetV4 module, and / or, replacing a plurality of Bottleneck modules in the C2f module in the neck network of the YOLOv8n network with GSBottleneck modules, and / or, adding an EMA module before the pooling layer at the end of the backbone network of the YOLOv8n network.

4. The pedestrian object detection method of claim 2, wherein, The pedestrian target detection model is trained, specifically including: inputting the historical denoising fusion image into the pedestrian target detection model based on the YOLOv8 model to obtain a predicted bounding box corresponding to the pedestrian target; determining a second loss according to the preset real bounding box and the predicted bounding box of the pedestrian target by the following formula, and training the pedestrian target detection model with the optimization target of minimizing the second loss: ， ； wherein, is a second loss, is a coordinate box of the real bounding box, ( , ) is a left-top corner coordinate of the bounding box, , ) is a right-bottom corner coordinate of the bounding box, is a coordinate box of the predicted bounding box, ( , ) is a left-top corner coordinate of the predicted box, , ) is a right-bottom corner coordinate of the predicted box, is the real bounding box, is the predicted bounding box, is a minimum Euclidean distance of a diagonal vertex pair of the real bounding box and the predicted bounding box, is a maximum Euclidean distance of a diagonal vertex pair of the real bounding box and the predicted bounding box, is a left-top corner coordinate of the real bounding box, is a right-bottom corner coordinate of the real bounding box, is a left-top corner coordinate of the predicted bounding box, is a left-top corner coordinate of the predicted bounding box.

5. The pedestrian object detection method of claim 1, wherein, The global image semantic features include scene structure features and illumination distribution features; The local edge texture features include image edge features and texture difference features.

6. A pedestrian object detection apparatus characterized by comprising: including: an acquisition module configured to acquire an infrared image and a visible light image to be detected; a forward noise adding module configured to input the infrared image and the visible light image to be detected into an improved diffusion model, and add noise in the infrared image and the visible light image to be detected through a forward process of the improved diffusion model to obtain an infrared noise image and a visible light noise image; a reverse denoising fusion module configured to extract global image semantic features and local edge texture features of the infrared noise image and global image semantic features and local edge texture features of the visible light noise image through a reverse process of the improved diffusion model in multiple levels and fuse them layer by layer to obtain a denoising fusion image; a detection module configured to input the denoising fusion image into a pedestrian target detection model to obtain a pedestrian target detection result; The reverse part of the improved diffusion model includes a denoising feature extraction layer and a feature fusion layer; The denoising feature extraction layer includes a first branch and a second branch in parallel, and the first branch and the second branch each include a plurality of serially connected comprehensive extraction modules, the first branch is configured to extract global image semantic features and local edge texture features of the visible light noise image through a plurality of serially connected comprehensive extraction modules in multiple levels, and the second branch is configured to extract global image semantic features and local edge texture features of the infrared noise image through a plurality of serially connected comprehensive extraction modules in multiple levels; Each comprehensive extraction module includes a reversible global capture module and a lightweight local capture module in parallel; the reversible global capture module includes a reversible neural network model, and the lightweight local capture module includes a Lite Transformer model; The feature fusion layer includes a plurality of serially connected Fusion layer modules. The Fusion layer module connected with the last integrated extraction module in the first branch and the second branch is taken as the first Fusion layer module; The first Fusion layer module is used for receiving the global image semantic features and the local edge texture features of the infrared noise image and the global image semantic features and the local edge texture features of the visible light noise image obtained by the last integrated extraction module in the first branch and the second branch, performing feature fusion to obtain an initial denoising fusion image, and transmitting the initial denoising fusion image to the next Fusion layer module, so that the next Fusion layer module performs feature fusion according to the initial denoising fusion image and the features obtained by the second last integrated extraction module in the first branch and the second branch, and the Fusion layer modules and the integrated extraction modules in the first branch and the second branch form a denoising fusion structure in reverse order.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the method in any one of claims 1-5.

8. A computer device, comprising: The computer program is stored in the memory and can be run on the processor, and the processor implements the method in any one of claims 1-5 when executing the program.

Images