Foreign matter intrusion detection method for low-illumination heavy haul railway

Abstract

The invention discloses a foreign matter intrusion detection method for a low-illumination heavy haul railway, and belongs to the field of foreign matter intrusion detection for railway security and protection, and the method comprises the following steps: collecting monitoring image data under different illumination conditions along the railway, marking the foreign matter type and position information of the monitoring image data, and obtaining a preprocessed image; performing image enhancement on the preprocessed image based on a Laplacian pyramid decomposition method and a low-frequency feature enhancement module to obtain an enhanced feature map containing multi-scale low-frequency information; and adopting a YOLOv10 model to carry out target detection on the enhanced feature map to obtain a foreign matter intrusion detection result. According to the invention, the detection performance of the railway safety monitoring system under the complex illumination condition is obviously improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of foreign object intrusion detection of railway security, and particularly relates to a low-illumination heavy-load railway foreign object intrusion detection method. BACKGROUND

[0002] Current foreign object intrusion detection technologies are mainly divided into two categories: contact type and non-contact type. Contact type detection technologies usually include fiber grating sensing and power grid sensing methods, which rely on a large number of equipment installations, thus having problems such as high cost, difficult maintenance, and difficulty in achieving continuous detection due to terrain limitations, and are currently mainly applied to foreign object detection such as tunnel mouth rockfall. Non-contact detection technologies include radar, ultrasonic, and infrared detection technologies based on physical sensing, as well as video intelligent analysis technologies based on image processing. In recent years, with the rapid development of computer vision and deep learning technologies, foreign object intrusion detection technologies based on intelligent analysis of monitoring videos have been developed and widely applied in the field of railway security. Compared with other technologies, video intelligent analysis has lower requirements for hardware, and can usually directly utilize existing video monitoring systems and embed algorithm modules to achieve detection functions.

[0003] However, in railway foreign object intrusion detection, although the existing YOLO series network models perform excellently in real-time detection, they still have obvious limitations in dealing with some special scenarios, especially in low-illumination environments. The YOLO network model extracts features of different scales through multiple convolution layers, and in low-illumination environments, detailed information may be lost due to the decrease in image brightness and contrast, making it difficult for the model to effectively distinguish targets from backgrounds, especially for smaller foreign object targets, the detection accuracy will be significantly reduced. This situation is particularly prominent in railway environments, not only limited to night scenes, but also including scenes with uneven light such as tunnels and under bridges, where the intruding objects may only have faint outlines or be obscured by strong shadows, making it difficult for the model to accurately identify and locate these objects.

[0004] In addition, the YOLO series network models rely on a large amount of labeled clear high-quality data for training, while most railway monitoring data has poor image quality under low-illumination conditions, and the lack of data sets directly affects the adaptability of the model. Although the latest YOLO versions attempt to introduce deeper feature extraction modules, attention mechanisms, and pyramid structures to enhance the network's feature expression ability, in low-light conditions, the noise of the image significantly increases, resulting in insufficient detection ability of the model for small targets and distant objects, which seriously affects the model accuracy and further challenges the practicality and reliability of the railway security system. Therefore, the present application proposes a low-illumination heavy-load railway foreign object intrusion detection method. SUMMARY

[0005] To solve the above technical problems, the present application provides a low-illumination heavy haul railway foreign matter intrusion detection method to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides a low-illumination heavy haul railway foreign matter intrusion detection method, comprising:

[0007] Collecting monitoring image data under different illumination conditions along the railway and labeling the foreign matter category and position information of the monitoring image data to obtain a preprocessed image;

[0008] Based on the Laplacian pyramid decomposition method and the low-frequency feature enhancement module, the preprocessed image is image-enhanced to obtain an enhanced feature map containing multi-scale low-frequency information;

[0009] The enhanced feature map is detected by a YOLOv10 model to obtain a foreign matter intrusion detection result.

[0010] Optionally, the process of image-enhancing the preprocessed image based on the Laplacian pyramid decomposition and the low-frequency feature enhancement module to obtain an enhanced feature map containing multi-scale low-frequency information comprises:

[0011] The preprocessed image is subjected to Gaussian blur and smoothing processing to remove high-frequency noise to obtain a denoised image;

[0012] The denoised image is decomposed based on the Laplacian pyramid decomposition method to output several low-frequency component feature maps;

[0013] The low-frequency component feature maps are up-sampled by a low-frequency feature enhancement module to obtain an enhanced feature map containing multi-scale low-frequency information.

[0014] Optionally, the process of decomposing the denoised image based on the Laplacian pyramid decomposition method comprises:

[0015] A Gaussian pyramid is constructed based on down-sampling operation and Gaussian filtering;

[0016] Based on the Gaussian filtering pyramid, a Laplacian pyramid is constructed by difference operation;

[0017] The denoised image is down-sampled based on the Laplacian pyramid to obtain several low-frequency component feature maps; wherein the resolution of the several low-frequency component feature maps is halved successively.

[0018] Optionally, the expression of the Laplacian pyramid is:

[0019] L l =G l -↑2(G l+1 )

[0020] In the formula, L l is the lth layer of the Laplacian pyramid, G l is the lth layer of the Gaussian pyramid, G l+1 is the l+1th layer of the Gaussian pyramid, and ↑2 is a bilinear up-sampling operation with an up-sampling factor of 2.

[0021] Optionally, the process of up-sampling the plurality of low-frequency component feature maps by using a low-frequency feature enhancement module to obtain an enhanced feature map containing multi-scale low-frequency information includes:

[0022] performing channel separation on the input low-frequency component feature map to generate a plurality of sub-feature maps;

[0023] performing adaptive average pooling operations of different scales on each sub-feature map respectively to extract multi-scale low-frequency features and obtain a pooling result of each scale;

[0024] performing up-sampling operations on the pooling result of each scale to restore the original resolution of the low-frequency component feature map to obtain an up-sampled feature map;

[0025] concatenating the up-sampled feature map in the channel dimension and performing convolution fusion on the concatenated feature by using a 1x1 convolution to generate an enhanced feature map containing multi-scale low-frequency information.

[0026] Optionally, the YOLOv10 model includes a backbone network, a neck, and a detection head.

[0027] The process of performing target detection on the enhanced feature map by using the YOLOv10 model includes:

[0028] inputting the enhanced feature map containing multi-scale low-frequency information into the backbone network to perform feature extraction and obtain multi-scale feature representations;

[0029] performing feature fusion on the multi-scale feature representations by using the neck network to obtain cross-fusion features;

[0030] based on the cross-fusion features, the detection head outputs an alien intrusion detection result.

[0031] Optionally, the process of inputting the enhanced feature map containing multi-scale low-frequency information into the backbone network to perform feature extraction and obtain multi-scale feature representations includes:

[0032] based on the enhanced feature map, multi-level feature representations are obtained by using multi-level convolution processing;

[0033] the multi-level feature representations are processed by using a cross-feature fusion method to obtain cross-layer connection features;

[0034] The cross-layer connection feature is sequentially processed by a spatial pyramid pooling module and a multi-head attention feature enhancement module to obtain a multi-scale feature representation.

[0035] Optionally, the detection head adopts a multi-scale detection head design, including:

[0036] The first detection head processes a feature map with a resolution of 80*80, and is used for detecting small-size targets.

[0037] The second detection head processes a feature map with a resolution of 40*40, and is used for detecting medium-size targets.

[0038] The third detection head processes a feature map with a resolution of 20*20, and is used for detecting large-size targets.

[0039] Compared with the prior art, the present application has the following advantages and technical effects:

[0040] The low-illumination heavy-load railway foreign matter intrusion detection method provided by the present application combines innovative image enhancement technology with an improved YOLOv10 model, significantly improving the detection performance of the railway safety monitoring system under complex lighting conditions. The method first uses Laplacian pyramid decomposition to extract multi-scale features from the input image, and cooperates with a specially designed low-frequency feature enhancement module to strengthen the structural information of the image, effectively solving the problem of serious detail loss in the low-illumination environment. The improved YOLOv10 model enhances the network's ability to extract multi-scale features through a cross-feature fusion module and a multi-head attention mechanism, especially significantly improving the detection accuracy of small targets and distant foreign matters. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application illustrated in the drawings and their descriptions are used to explain the present application and are not intended to limit the present application. In the drawings:

[0042] Figure 1 It is an image enhancement framework structure diagram of the embodiment of the present application;

[0043] Figure 2 It is a structure diagram of the low-frequency feature enhancement module of the embodiment of the present application;

[0044] Figure 3 It is a structure diagram of the improved YOLOv10 network of the embodiment of the present application;

[0045] Figure 4 It is a cross-feature fusion module diagram of the embodiment of the present application;

[0046] Figure 5 It is a multi-head attention feature enhancement module diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0047] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0049] The present scheme provides a target detection method based on an improved YOLOv10 model, which is suitable for the application scenario of low-illumination heavy-load railway foreign object intrusion detection, and aims to solve the problem of insufficient performance of existing target detection technology under low-illumination conditions. The method comprises: collecting normal-illumination and low-illumination rail foreign object image data, labeling and dividing into a training set and a validation set; constructing an image enhancement framework combining a Laplacian pyramid frequency domain decomposition and a multi-branch enhancement extraction module of "low-frequency feature enhancement + residual unit" for feature extraction and image enhancement; loading the enhanced image into the YOLOv10 model for training, iterative optimization to obtain the final foreign object intrusion detection model; in the inference stage, the model detects the target and outputs the position information and the class prediction result. The present scheme effectively enhances the detection performance of the model in the low-illumination environment, improves the overall detection precision and robustness.

[0050] The present application solves the following technical problems: (1) Improve the detection accuracy in low-light scenes. Railway foreign object intrusion detection faces complex lighting conditions, especially at night, in shadow or strong light reflection environment, the existing detection technology often causes the target boundary to be blurred due to uneven lighting, increasing false positives and false negatives. The present application designs an image enhancement framework that preserves the multi-scale detail information of the image and enhances the overall structural features: the image is decomposed into different scales through multi-branch frequency domain decomposition, preserving the high-frequency detail levels; the low-frequency feature enhancement module enhances the global information and contour of the image, so that the enhanced image can provide more detailed information and clearer target contour in complex scenes. The processed enhanced image is input into the YOLOv10 target detection model, which can effectively improve the performance of the model in low-light scenes. (2) Improve the accuracy of small target detection. One of the difficulties of railway foreign object detection is to accurately identify small foreign objects at a distance, especially in complex backgrounds, small target foreign objects are easily ignored in traditional detection methods. The present application combines YOLOv10 and image enhancement technology, through frequency domain decomposition and multi-branch enhancement extraction module of "low-frequency feature enhancement + residual unit", extracts the low-frequency information in the image, enhances the global contour and structural features, so that the model can maintain clarity and stability on low-resolution feature maps when extracting small target features. The low-frequency feature enhancement module designed in this scheme can effectively reduce the interference background and highlight the small target features, thereby significantly improving the detection accuracy of small targets in long distance and complex background, meeting the needs of railway foreign object detection.

[0051] As shown in Figure 1 In this embodiment, a low-light heavy-load railway foreign object intrusion detection method is provided, comprising the following steps: collecting monitoring image data under different lighting conditions along the railway and labeling the foreign object category and position information of the monitoring image data to obtain preprocessed images; based on the Laplacian pyramid decomposition method and the low-frequency feature enhancement module, the preprocessed images are enhanced to obtain enhanced feature maps containing multi-scale low-frequency information; and the YOLOv10 model is used to detect the enhanced feature maps to obtain foreign object intrusion detection results.

[0052] S1, data collection and labeling. High-definition camera equipment is used to obtain foreign object image data under different lighting conditions of heavy-load railways, and the images are accurately labeled to form training and verification data sets, ensuring that the model can adapt to multiple environments.

[0053] This embodiment aims to meet the needs of foreign object intrusion detection in heavy-load railway application scenarios. First, data collection and labeling are performed. High-definition camera equipment installed along the heavy-load railway monitors and photographs the rail and its surrounding area for a long time to obtain image data under different lighting conditions. The collection time covers various situations such as daytime, nighttime and bad weather, with a focus on data collection at night or in other low-light environments.

[0054] The collected picture data is labeled using professional tools such as labelme. The main labeling content needs to include: the classification labeling of various foreign matters on the track (such as people, falling stones, coal cinders, scattered equipment, etc.), and the labeling box is accurate to the edge of the foreign matter to ensure the labeling accuracy. The labeling information of each image includes the position coordinates of the foreign matter, the category label, and whether it is in low-illumination or normal light conditions, etc., to obtain the preprocessed image.

[0055] S2, image enhancement. The preprocessed image is subjected to multi-scale decomposition and enhancement by using Laplacian pyramid decomposition and low-frequency feature enhancement module, the high-frequency details are retained and the low-frequency structure is strengthened, the image quality under low-illumination condition is improved, and more ideal image input is provided for subsequent detection.

[0056] Further, the process of obtaining the enhanced feature map containing multi-scale low-frequency information based on the preprocessed image includes: the preprocessed image is subjected to Gaussian blur and smoothing processing, high-frequency noise is removed to obtain a denoising image; the denoising image is decomposed based on the Laplacian pyramid decomposition method, and a plurality of low-frequency component feature maps are output; the low-frequency feature enhancement module is used to up-sample the plurality of low-frequency component feature maps, and an enhanced feature map containing multi-scale low-frequency information is obtained.

[0057] The embodiment designs a multi-branch enhancement extraction framework combining frequency domain decomposition and "low-frequency feature enhancement + residual unit", which performs multi-scale detail and global structure enhancement processing on the input image, to specifically improve the feature expression ability of the image under low-illumination condition, so as to provide better input image for subsequent target detection task.

[0058] This step decomposes the input preprocessed image into feature maps of different scales, and performs residual feature extraction and low-frequency enhancement processing on each scale. Finally, through multi-level up-sampling and feature fusion, an enhanced image with more rich details and global features is reconstructed, as shown in Figure 1 .

[0059] Multi-scale decomposition of the image is the basis of the enhancement process. First, the preprocessed image is subjected to Gaussian blur and smoothing processing to remove high-frequency noise and obtain a denoising image; then, Laplacian down-sampling operation is performed to reduce the resolution, and a plurality of low-frequency component feature maps are output. The process of Laplacian pyramid decomposition decomposes the image into L=4 layers of features, each layer retaining different frequency band detail information. Let the input image be First, the Gaussian pyramid is constructed Each layer of the Gaussian pyramid is generated by down-sampling operation and Gaussian filtering, as shown in formula (1), where ↓2 represents a bilinear down-sampling operation with a down-sampling factor of 2, K σ represents the Gaussian kernel.

[0060] Gl = ↓2(G l-1 = Kσ) (1)

[0061] Subsequently, the Laplacian pyramid is constructed by difference operation The resolution is restored by bilinear interpolation, as shown in equation (2), and the frequency domain of the feature map gradually increases from left to right.

[0062] L l = G l - ↑2(G l+1 ) (2)

[0063] The input image size is set to 640x640, and after a series of downsampling operations, a plurality of feature maps with resolution halved in turn are generated, namely feature map 0 (640x640), feature map 1 (320x320), feature map 2 (160x160) and feature map 3 (80x80). Each layer of the processed feature map retains the overall structure and low-frequency information of the original image, while the high-frequency details are compressed to lower resolutions layer by layer. At each scale, the low-resolution image of the current scale is restored to the resolution of the previous scale by upsampling operation, and the difference with the original image of the previous scale is retained, which retains the detail information lost by blurring or downsampling at each scale, so that the high-frequency details can be independently extracted and subsequently enhanced. Figure 1 Figure 2 Figure 3

[0064] The low-frequency feature enhancement module is mainly used to strengthen the low-frequency information in the image, and its structure is shown in Figure 2 Low-frequency information usually contains the outline, shape and basic structure of an object, which is crucial for the stability and robustness of foreign object detection in complex environments.

[0065] Further, the process of upsampling a plurality of low-frequency component feature maps by using the low-frequency feature enhancement module to obtain an enhanced feature map containing multi-scale low-frequency information includes: performing channel separation on the input low-frequency component feature map to generate a plurality of sub-feature maps; performing adaptive average pooling operation of different scales on each sub-feature map respectively to extract multi-scale low-frequency features and obtain the pooling result of each scale; restoring each scale of the pooling result to the original resolution of the low-frequency component feature map by upsampling operation to obtain an upsampled feature map; concatenating the upsampled feature maps in the channel dimension, and performing convolution fusion on the concatenated features by 1x1 convolution to generate an enhanced feature map containing multi-scale low-frequency information.

[0066] ​​​The low-frequency feature enhancement module is used to up-sample a plurality of low-frequency component feature maps to obtain an enhanced feature map containing multi-scale low-frequency information. The low-frequency feature enhancement module first performs a 1x1 convolution operation on the input low-frequency component feature map to adjust the channel number and preliminarily extract the feature; then performs a channel separation operation to generate a plurality of sub-feature maps, each sub-channel passes through an adaptive average pooling of different scales, only allows part of the cutoff frequency to pass, enhances the composition of low-frequency information, and obtains the pooling result of each scale. For the pooling result of each scale, further up-sampling operation is performed to restore it to the original resolution of the input feature map to obtain the up-sampled feature map. The up-sampled feature maps are spliced in the channel dimension and fused through 1x1 convolution to generate an enhanced feature map containing multi-scale low-frequency information. This operation ensures that the overall contour information of the image in the low-light environment will not be lost due to insufficient light.

[0067] S3, YOLOv10 model training and target detection. Cross feature fusion module (C2f) and multi-head attention module are introduced into the skeleton network to enhance the feature extraction capability, and through multi-scale feature fusion and iterative optimization, the precise detection of small targets and distant foreign objects is realized.

[0068] The YOLOv10 network model includes a backbone network (Backbone), a neck (Neck) and a detection head (Head) to realize multi-scale high-precision target detection, and the network structure is as shown in Figure 3 .

[0069] Further, the process of using the YOLOv10 model to detect the target of the enhanced feature map includes: inputting the enhanced feature map containing multi-scale low-frequency information into the skeleton network to extract features to obtain multi-scale feature representation; the multi-scale feature representation is fused by the neck network to obtain cross-fusion features; based on the cross-fusion features, the detection head outputs the foreign object intrusion detection result.

[0070] Further, the process of inputting the enhanced feature map containing multi-scale low-frequency information into the skeleton network to extract features to obtain multi-scale feature representation includes: based on the enhanced feature map, using multi-level convolution processing to obtain multi-level feature representation; using a cross feature fusion method to process the multi-level feature representation to obtain cross-layer connection features; the cross-layer connection features are processed in turn by a spatial pyramid pooling module and a multi-head attention feature enhancement module to obtain multi-scale feature representation.

[0071] The main goal of the model training stage is to extract and fuse multi-layer features from the dataset images, and iteratively optimize according to the loss function. First, the skeleton network of the model serves as the feature extraction part, which processes the input image layer by layer to obtain multi-scale feature representation. The input image is gradually expanded from the initial RGB 3-channel image to 64, 128, and 256-channel high-level features through a series of 3x3 convolution layers, and finally to multi-level feature representation. The cross-feature fusion module specific to YOLOv10 establishes cross-layer connections between feature maps at different levels, preserves detailed information through group compression downsampling and residual connection. The end of the skeleton network is composed of a spatial pyramid pooling module and a multi-head attention feature enhancement module. The former is used to extract global context information, and the latter enhances the expression of important features by adaptively extracting different regions of the image. Finally, multi-scale feature representation is obtained. The following is a specific introduction to the key modules:

[0072] The cross-feature fusion module is composed of 1x1 convolution, feature splitting, residual unit, and concatenation components, as shown in Figure 4 . The input feature map is first adjusted in channel number by 1x1 convolution to reduce redundant information; then the feature map is split, and each sub-feature map enters an independent residual unit for processing to extract different scale details and semantic information, enriching the diversity of features; then the outputs of each residual unit are concatenated in the channel dimension to form a fusion feature map containing multi-level features. Finally, the concatenated features are integrated by 1x1 convolution. This design can effectively reduce the computational burden while preserving multi-level information and improving the performance of the target detection model in multi-scale target detection.

[0073] The multi-head attention feature enhancement module is a parallel spatial attention mechanism that improves the expression ability of the feature map and the learning ability of the attention weight through an effective multi-branch structure, as shown in Figure 5 . The input feature map is adjusted in channel number by 1x1 convolution, separated into parallel channels, and processed. Each sub-channel enters the multi-head attention mechanism independently to capture spatial context information on different channels, thereby enhancing the model's attention to important feature regions. The output results are added element by element with the original feature map to fuse the features, and then added with the features output by the feedforward layer to further integrate the spatial context information from the multi-head attention. Finally, the output features of all sub-channels are concatenated in the channel dimension to form a feature map, which is then compressed in channel number by 1x1 convolution. This module improves the target recognition ability of the model through multi-level attention capture and feature fusion.

[0074] After the backbone network outputs multi-scale feature representations, they are input into the neck network for further fusion of features of different scales to obtain cross-fusion features to adapt to the requirement of multi-scale target detection. Each layer after cross-fusion of features is up-sampled to restore it to a higher resolution through bilinear interpolation, and is spliced with a feature map of a previous layer to form a new feature representation. This step effectively combines shallow features and deep features to realize multi-scale feature fusion and ensure that small targets and large targets are effectively expressed in the detection process.

[0075] The detection head is the final output part of YOLOv10 and is responsible for converting the features extracted and fused in the foregoing into position information of a detection frame and a class prediction result. The detection head adopts a multi-scale detection head design and detects small targets (80x80), medium targets (40x40) and large targets (20x20) respectively to adapt to the requirement of detection of targets of different sizes. Each scale of the detection head includes a cross-feature fusion module, a convolution module and a splicing module to ensure that spatial and semantic information is fully utilized at each scale. For each scale, the detection head outputs a bounding box coordinate (center of mass xy, width and height wh) of each detection frame, a confidence score c and a target class p.

[0076] In the inference stage, a plurality of operations are combined into a unified operation by using a block compression technique to improve inference efficiency and reduce calculation and memory overhead. The operation can be represented by a symbol , and the formula is as shown in (3): when M branches are combined into a convolution kernel, wherein χ is an input feature map, W j is a convolution kernel weight of the jth branch, o and i respectively represent the number of output and input channels, τ is a channel dimension weighting coefficient (vector) learned in training and is used to automatically adjust the importance of outputs of the branches, y is an output feature map, and the weighted sum of results of the M branches is obtained.

[0077]

[0078] The loss function of the application mainly includes three parts: a positioning loss L loc , a confidence loss L conf and a classification loss L cls , wherein (x i , y i , h i , w i , c i ) is information of an ith prediction frame, including a center point horizontal and vertical coordinate, a width and a height and a confidence, is information of a corresponding real frame, is a [0, 1] binary parameter, 1 for a target and 0 for no target. These loss terms jointly act, the total loss L adjusts model parameters through back propagation, iteratively optimizes model performance, and the formula of the loss function is as follows:

[0079]

[0080] Based on the above technical solutions, the present application has the following technical effects:

[0081] (1) Significantly improve the detection performance under low illumination conditions. The image enhancement module of the present application can significantly improve the performance of the YOLOv10 target detection model under low illumination conditions. By extracting multi-scale high-frequency and low-frequency features, the global structure and details of the image are effectively enhanced, which can provide YOLOv10 with clearer and higher contrast image input, thereby greatly improving the detection accuracy of the model under low illumination conditions, and further improving the reliability and safety of the overall detection.

[0082] (2) Improve the precision and stability of small target detection. Small foreign objects may pose a significant threat to the safe operation of trains. The image enhancement module of the present application enhances the information retention of small targets at multiple levels through multi-scale feature extraction and fusion, enabling the YOLOv10 network to identify small size targets in high resolution feature maps, and avoiding the loss of small targets due to scale changes.

[0083] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A low-illumination heavy-load railway foreign object intrusion detection method, characterized by, The method comprises the following steps: Collecting monitoring image data under different light conditions along the railway and labeling foreign matter categories and position information of the monitoring image data to obtain preprocessed images; Performing image enhancement on the preprocessed images based on a Laplacian pyramid decomposition method and a low-frequency feature enhancement module to obtain enhanced feature maps containing multi-scale low-frequency information; Performing target detection on the enhanced feature maps by using a YOLOv10 model to obtain foreign matter intrusion detection results.

2. The low-illumination heavy haul railway foreign object intrusion detection method according to claim 1, characterized in that, The process of performing image enhancement on the preprocessed images based on the Laplacian pyramid decomposition and the low-frequency feature enhancement module to obtain enhanced feature maps containing multi-scale low-frequency information comprises: The preprocessed images are subjected to Gaussian blur and smoothing processing to remove high-frequency noise to obtain denoised images; The denoised images are decomposed based on the Laplacian pyramid decomposition method to output a plurality of low-frequency component feature maps; The plurality of low-frequency component feature maps are up-sampled by using a low-frequency feature enhancement module to obtain enhanced feature maps containing multi-scale low-frequency information.

3. The low-light heavy haul rail foreign object intrusion detection method of claim 2, wherein, The process of decomposing the denoised images based on the Laplacian pyramid decomposition method comprises: A Gaussian pyramid is constructed based on a down-sampling operation and a Gaussian filter; A Laplacian pyramid is constructed based on the Gaussian filter pyramid by using a difference operation; The denoised images are down-sampled based on the Laplacian pyramid to obtain a plurality of low-frequency component feature maps; wherein the resolutions of the plurality of low-frequency component feature maps are halved successively.

4. The low-light heavy haul rail foreign object intrusion detection method of claim 3, wherein, The expression of the Laplacian pyramid is: L l = G l -↑2(G l+1 ) where L l is the l-th layer of the Laplacian pyramid, G l is the l-th layer of the Gaussian pyramid, G l+1 is the (l+1)-th layer of the Gaussian pyramid, and ↑2is a bilinear upsampling operation with an upsampling factor of 2.

5. The low-light heavy haul rail foreign object intrusion detection method of claim 3, wherein, The process of up-sampling the plurality of low-frequency component feature maps by using a low-frequency feature enhancement module to obtain enhanced feature maps containing multi-scale low-frequency information comprises: Channel separation is performed on the input low-frequency component feature maps to generate a plurality of sub-feature maps; An adaptive average pooling operation of different scales is performed on each sub-feature map to extract multi-scale low-frequency features and obtain a pooling result of each scale; Each scale of the pooling result is up-sampled to restore the original resolution of the low-frequency component feature map to obtain an up-sampled feature map; The up-sampled feature maps are spliced in the channel dimension, and the spliced features are convolved and fused by using a 1×1 convolution to generate enhanced feature maps containing multi-scale low-frequency information.

6. The low-light heavy haul rail foreign object intrusion detection method of claim 1, wherein, The YOLOv10 model comprises: a skeleton network, a neck, and a detection head; The process of performing target detection on the enhanced feature maps by using the YOLOv10 model comprises: The enhanced feature maps containing multi-scale low-frequency information are input into the skeleton network for feature extraction to obtain multi-scale feature representations; The multi-scale feature representations are fused by the neck network to obtain cross-fusion features; Based on the cross-fusion features, the detection head outputs foreign matter intrusion detection results.

7. The low-light heavy haul rail foreign object intrusion detection method of claim 6, wherein, The process of inputting the enhanced feature maps containing multi-scale low-frequency information into the skeleton network for feature extraction to obtain multi-scale feature representations comprises: Based on the enhanced feature maps, multi-level feature representations are obtained by using multi-level convolution processing; The multi-level feature representations are processed by using a cross-feature fusion method to obtain cross-layer connection features; The cross-layer connection feature is processed by a spatial pyramid pooling module and a multi-head attention feature enhancement module in sequence to obtain a multi-scale feature representation.

8. The low-light heavy haul rail foreign object intrusion detection method of claim 7, wherein, The detection head adopts a multi-scale detection head design, including: A first detection head processes a feature map with a resolution of 80x80, for detecting small-size targets; A second detection head processes a feature map with a resolution of 40x40, for detecting medium-size targets; A third detection head processes a feature map with a resolution of 20x20, for detecting large-size targets.

Images