Track turnout foreign matter detection and type analysis method and system

Abstract

The invention provides a rail turnout foreign matter detection and type analysis method and system, and the method comprises the steps: arranging a first camera group at the front end of a train, recognizing the position of a turnout through a preset target detection model, and calibrating foreign matters in the turnout through a preset foreign matter detection model; shooting an image calibrated to the foreign matter in the turnout through a second camera group as a to-be-analyzed image; extracting a second boundary region rolled by a wheel on the foreign matter from the to-be-analyzed image through a preset wheel rolling track detection model, performing coarse classification on the foreign matter in the to-be-analyzed image to obtain a coarse classification result, and selecting a boundary detection strategy to identify the first boundary region according to the coarse classification result in the foreign matter image; and respectively extracting corresponding texture features in the first boundary region and the second boundary region through a deep learning network, performing one-to-one mapping on the extracted texture features to obtain a mapping relation group, performing analysis through the mapping relation group to obtain a feature change rate, and performing analysis through a multi-feature analysis strategy to obtain a specific foreign matter type.

Description

Technical Field

[0001] The present invention relates to an orbit foreign object recognition system, and in particular to a method and system for detecting and analyzing the type of foreign objects on a railway switch Background Art

[0002] In the railway transportation system, as a key connecting device of the track line, the normal operation of the switch plays a decisive role in the safe and efficient running of trains. Once foreign objects appear in the switch area, it may lead to serious accidents such as train derailment and signal failure, causing huge economic losses and potential safety hazards.

[0003] At present, the detection of foreign objects on railway switches mainly adopts two methods: manual inspection and traditional sensor detection. Manual inspection relies on staff to visually inspect the switch area regularly. It not only has low efficiency and high labor intensity, but also is affected by human factors, making it difficult to detect tiny foreign objects or conduct accurate detection under bad weather conditions. Traditional sensor detection, such as infrared sensors and pressure sensors, although can achieve automated detection to a certain extent, these sensors can often only detect the presence of foreign objects and cannot accurately analyze the types of foreign objects. Different types of foreign objects have different degrees of impact on the operation of the switch. For example, metal foreign objects may cause a short circuit in the circuit, while foreign objects such as paper towels and plastic bags may affect the normal rotation of the switch. Therefore, it is difficult to formulate targeted treatment measures without analyzing the types of foreign objects, posing a relatively high safety risk. In addition, some existing image-based detection methods usually only use a single camera to collect images. Due to the single perspective and limited information, it is easy to miss detections and misdetections in complex environments, and cannot meet the requirements of high-precision and high-reliability foreign object detection on railway switches. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a method and system for detecting and analyzing the type of foreign objects on a railway switch to overcome the above-mentioned defects in the existing technology.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: A method for detecting and analyzing the category of foreign objects on a railway switch, comprising: A step of identifying foreign objects on the switch. A first camera group is provided at the front end of the train. The first camera group is used to shoot a track video, and frame extraction detection is performed. Then, the position of the switch is identified through a preset target detection model, and foreign objects in the switch are calibrated through a preset foreign object detection model. The corresponding image in the video is intercepted as a foreign object image; An image acquisition step. The second camera group provided behind the train wheels is used to shoot the image of the foreign object calibrated in the switch as an image to be analyzed; Image processing step: Use a preset wheel rolling trajectory detection model to extract the second boundary region of foreign objects on the wheel in the image to be analyzed, and perform a rough classification on the foreign objects in the image to be analyzed to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects. Select a boundary detection strategy according to the rough classification result in the foreign object image to identify the boundary region as the first boundary region; Detail feature extraction step: Use a deep learning network to extract corresponding texture features in the first boundary region and the second boundary region respectively, map the extracted texture features one by one to obtain a mapping relationship group, and then analyze the feature change rate through the mapping relationship group. The feature change rate includes texture contrast ratio, color contrast ratio and shape contrast ratio. Then, analyze the specific foreign object type according to the texture contrast ratio, color contrast ratio and shape contrast ratio through a multi-feature analysis strategy.

[0006] Preferably, the flexible foreign objects include paper towels, plastic bags, and fabrics. When the rough classification result is a flexible foreign object, output a probability distribution vector representing the probability values of each specific category in the flexible foreign object. The database is set with foreign object type labels, corresponding priority boundary recognition strategies for each category, and standard sample features. Query the database according to the distribution vector of the probability values of each specific category in the rough classification result to obtain the corresponding boundary detection strategy.

[0007] Preferably, when the probability values of multiple specific categories all reach the threshold, obtain the first boundary region corresponding to each category of foreign objects according to the recognition strategy of each category respectively, analyze the foreign object boundary in the foreign object image through an image comparison strategy, determine the first boundary region according to the image similarity, and adjust the probability value according to the comparison result.

[0008] Preferably, the boundary detection strategy includes: Targeted image processing step: According to the rough classification result, when the foreign object is a paper towel or a fabric, enhance the image contrast and retain the edges, and use multi-scale edge detection combined with morphological operations to strengthen the edges; when the foreign object is a plastic bag, suppress the reflection through a polarization strategy, and use a deblurring model based on deep learning and difference repair to repair the boundary; Model training and inference step: Dynamically adjust the loss weight by combining foreign object type specificity and fuzzy IoU loss. The foreign object type specificity loss includes texture constraint loss for paper towels and fabrics and transparent region compensation loss and reflection region suppression loss for plastic bags. The fuzzy IoU loss is calculated based on fuzzy set theory to calculate the fuzzy intersection and construct.

[0009] Preferably, a position recognition step is further included. The position of the foreign object is judged according to the contour lines of the first boundary region and the track turnout, and the deformation trend of the foreign object is analyzed according to the contour lines of the second boundary region and the first boundary region. The position of the deformed foreign object is analyzed, and a danger warning is generated according to the type of the foreign object.

[0010] Preferably, the multi-feature analysis strategy includes combining the feature change rate with the probability weight of rough classification, and using a weighted voting mechanism to output the type of foreign object. When there is a conflict between the feature change rate and the rough classification result, the weights are dynamically adjusted according to the credibility of different features.

[0011] Preferably, a verification step is further included. A verification module is also provided on the train. The verification module is arranged behind the second camera group. The verification module includes a main camera, an auxiliary camera and a lidar. According to the type of the foreign object, the lighting intensity and camera parameters of the auxiliary camera are adjusted, and the image data collected by the main camera and the auxiliary camera and the laser electrical signals collected by the lidar are combined. The data is verified based on multi-modal data fusion processing, and an abnormal analysis result is obtained based on the verification result and compared with the type of the foreign object to judge whether they are consistent.

[0012] Preferably, a small target positioning strategy is set in the target detection model, including using a network architecture for feature extraction, and embedding an attention mechanism module in the network. The attention mechanism module includes a dynamic channel weight module, which assigns high weights to the key feature channels of the turnout data set through the SE-Net structure, and performs cross-scale channel interaction to screen the key features of the turnout target. And through the spatial attention module, the global position of the turnout is captured through global pooling to generate an attention map to locate the target area.

[0013] Preferably, the second camera group is a high-definition camera. When the first camera group detects the turnout, the second camera is adjusted so that the second camera captures an image of the area where the turnout is located.

[0014] A system for detecting and classifying foreign objects on a track turnout includes: A turnout foreign object recognition module. A first camera group is provided at the front end of the train. The first camera group is used to capture a track video, perform frame extraction detection, then identify the turnout position through a preset target detection model, and calibrate the foreign objects in the turnout through a preset foreign object detection model, and intercept the corresponding image in the video as a foreign object map; An image acquisition module, which captures an image of the foreign object calibrated in the turnout through a second camera group arranged behind the train wheels as an image to be analyzed; The image processing module extracts the second boundary area of the foreign object run over by the wheel in the image to be analyzed through a preset wheel rolling trajectory detection model, and roughly classifies the foreign object in the image to be analyzed to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects. According to the rough classification result in the foreign object image, a boundary detection strategy is selected to identify the boundary area as the first boundary area; The detailed feature extraction module extracts corresponding texture features in the first boundary area and the second boundary area respectively through a deep learning network, maps the extracted texture features one by one to obtain a mapping relationship group, and then analyzes the feature change rate through the mapping relationship group. The feature change rate includes texture contrast rate, color contrast rate and shape contrast rate. Then, according to the texture contrast rate, color contrast rate and shape contrast rate, a specific foreign object type is analyzed through a multi-feature analysis strategy.

[0015] The beneficial effects of the present invention: Through a multi-stage and multi-strategy processing method, making full use of the image information before and after the foreign object is pressed, combined with multi-feature analysis, various foreign objects can be accurately identified, and the false detection and missed detection rates can be effectively reduced. For foreign objects of different types and different shapes, especially flexible foreign objects, through a variety of preset boundary recognition strategies and a rich standard sample library, good detection and classification effects can be achieved, adapting to the complex and changeable track environment. Based on the high computing power of image processing and machine learning algorithms, foreign object detection and type analysis can be completed in real time during the train operation, providing timely guarantee for the train operation safety. Brief Description of the Drawings

[0016] Figure 1 is the overall flowchart of the present invention; Figure 2 is the image processing flowchart of the present invention; Figure 3 is the module connection diagram of the present invention; Figure 4 is the layout diagram of the camera group of the present invention. Detailed Embodiments

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0018] It should be noted that when a component is referred to as "fixed to" another component, it can be directly on the other component or there may also be an intermediate component. When a component is considered to be "connected to" another component, it can be directly connected to the other component or there may be an intermediate component at the same time. When a component is considered to be "disposed on" another component, it can be directly disposed on the other component or there may be an intermediate component at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0020] The following further details the embodiments of the present invention with reference to the accompanying drawings: As Figures 1-4 shown, the present invention provides a method for detecting foreign objects and analyzing their categories in a track turnout, including: The step of identifying foreign objects in the turnout: A first camera group is provided at the front end of the train. The first camera group is used to capture the track video, and frame extraction detection is performed. Then, the position of the turnout is identified through a preset target detection model, and foreign objects in the turnout are calibrated through a preset foreign object detection model. The corresponding image in the video is intercepted as a foreign object image. If a turnout exists in the current image frame, it enters the image acquisition step; if no turnout exists in the current image frame, the detection continues. A first camera group is provided at the front end of the train. This camera group is composed of multiple cameras at different angles and can capture the track video comprehensively and without dead angles. During the running of the train, the first camera group continuously captures the track video and performs frame extraction detection at a certain time interval (such as extracting 1-2 frames per second). A pre-trained target detection model is used to detect the position of the turnout in the extracted image frames. The target detection model adopts an algorithm based on deep learning, such as the YOLO series algorithm, and is trained through a large dataset containing turnout images, so that the model can learn the shape, structure and other features of the turnout. The position of the turnout is identified through the target detection model, and foreign objects in the turnout are calibrated through a preset foreign object detection model. According to the turnout position output by the target detection model, a detection ROI is generated, the ROI area is image-enhanced to improve the visibility of foreign objects under low light, and the foreign objects in the turnout are calibrated through the foreign object detection model, and the image is intercepted as a foreign object image.

[0021] Image acquisition step: The second camera group set behind the train wheels takes pictures of the foreign objects calibrated in the turnout as the images to be analyzed; when turnout foreign objects are detected, the current image containing the turnout is obtained as the foreign object image. At the same time, a second camera group is set behind the train wheels. This camera group consists of multiple high-definition cameras, and the positions and angles of the cameras are precisely calibrated. When the train wheels pass over the foreign object, the second camera group ensures that it can take pictures of the track area image at the same position as the foreign object image, and the taken picture is used as the image to be analyzed. By using two camera groups to take pictures of the same area at the same position, the differences of the foreign object after being run over by the wheels can be clearly compared. After comparative analysis of the differences between the two, a more accurate analysis of the foreign object type can be carried out.

[0022] The second camera group is a high-definition camera. When the first camera group detects the turnout, the second camera is adjusted so that the second camera takes pictures of the area where the foreign object in the turnout is located; the first camera is a camera with lower clarity. During the shooting process, the track can be continuously shot and stored. And the second camera is a high-definition camera. When the first camera detects the turnout foreign object, the second camera is started to take pictures of the turnout foreign object at the same position as the first camera. By using different cameras to take pictures of the turnout foreign object, during the process of ensuring continuous shooting and detection, the second camera can be used to take high-definition pictures of the foreign object, which is convenient for obtaining the surface texture features of the foreign object.

[0023] The small target localization strategy is set in the target detection model, including using the network architecture for feature extraction, embedding an attention mechanism module in the network. The attention mechanism module includes a dynamic channel weight module, which assigns high weights to the key feature channels of the turnout data set through the SE-Net structure and performs cross-scale channel interaction to screen the key features of the turnout target. And through the spatial attention module, it captures the global position of the turnout through global pooling and generates an attention map to locate the target area.

[0024] A multi-layer convolutional neural network is used as the basic framework, and the network is divided into three parts: the shallow layer, the middle layer, and the deep layer, gradually extracting turnout features from details to the whole: The shallow layer network focuses on retaining image details, uses small-sized convolutional kernels, such as 3×3 convolutional kernels, and sets a relatively large number of channels, such as 64 channels. In the first convolutional layer, a 3×3×3 convolutional kernel (assuming the input image is an RGB three-channel image) is used to perform a convolutional operation on the original image, and a feature map with 64 channels is output. This can capture basic features such as the fine edges of the switch point rail and the local structure of the switch machine, laying a foundation for extracting more advanced features later. The middle layer network increases the number of convolutional kernels, for example, from 64 to 128, and at the same time uses pooling operations, such as 2×2 max pooling, to expand the network's observation range of the image while reducing the resolution of the feature map. After two convolutional layers and one pooling operation, the size of the feature map will be reduced by half from the original size, and the number of channels will increase to 128. At this time, the network can initially understand the combination relationship and relative position between the various components of the turnout. The deep layer network introduces dilated convolution and global average pooling techniques. Taking dilated convolution with a dilation rate of 2 as an example, without increasing the number of network parameters, the network can "see" a larger range of image content, expanding the original 3×3 observation range to 5×5. Combining with global average pooling, the overall semantic information of the image is obtained, enabling the network to understand the position and role of the turnout in the entire track scene. A two-way feature fusion path is constructed. In the top-down path, the feature map with semantic information at the high level is upsampled to a size similar to that of the low-level feature map, and then added to the low-level high-resolution feature map; in the bottom-up path, the low-level feature map containing detailed information is passed to the high level through convolutional and pooling operations. For example, in the second layer of the FPN, the semantic feature map with a size of H / 8×W / 8 and 256 channels at the high level is upsampled to a size of H / 4×W / 4, and then added to the detailed feature map with the same size and 128 channels at the middle level to obtain the fused feature map, thus realizing the complementary advantages of turnout features at different scales and making the network's understanding of the turnout more comprehensive.

[0025] The dynamic channel weight module is based on the principle of SE-Net. By automatically adjusting the weights of each channel, it strengthens the key feature channels related to the turnout and realizes the channel information interaction between different scales. The SE-Net module includes the Squeeze stage: performing global average pooling operation on the feature map obtained by network feature extraction. And the Excitation stage: processing the compressed vector through two fully connected layers. First, the vector is mapped to a lower-dimensional space to reduce the number of parameters and introduce non-linear changes, and then it is mapped back to the original number of channels. Finally, through the activation function, the attention weight corresponding to each channel is obtained; enabling the network to learn to recognize the feature channels related to the key components of the turnout, such as the switch rail, frog, etc. During the training process, when an image containing the switch rail of the turnout is input, the network will gradually increase the weights of the channels that describe the features such as the shape and material texture of the switch rail. At the same time, a transfer mechanism for channel weights is established between different levels of the feature pyramid network: the key channel weights learned from the high-level feature map are fused into the corresponding channels of the low-level feature map by weighted summation. In this way, while the low-level feature map retains the original details, it can more prominently show the content related to the key information of the turnout. After the low-level feature map undergoes convolution and pooling operations, the local detail information it contains is fed back to the high-level feature map by adjusting the attention weights to help the high-level feature map better understand the overall structure of the turnout.

[0026] The spatial attention module realizes the accurate positioning of the target area of the turnout by integrating the global information of the image and generating an attention map, effectively reducing the interference of background information; performing global average pooling operations on the input feature map in the horizontal and vertical directions respectively: concatenating the feature vectors in these two directions along the channel dimension, and then processing them through a 7×7×1 convolutional layer to fuse and reduce the dimension of the features, and finally outputting an attention map with the same size as the input feature map; each pixel value in the attention map represents the possibility of the existence of the turnout target at that position. By multiplying the corresponding elements of the attention map and the original feature map, the adjustment of the feature map is realized. For small turnout targets, since global pooling synthesizes the information of the entire image, it will not be interfered by the surrounding background information because the small target occupies a small proportion in the image. For example, when there is a small target of the switch rail in the image, the attention map will assign a higher weight value to the area where the switch rail is located and a lower weight value to the background area. After the multiplication operation, the features in the switch rail area will be significantly enhanced, and the features in the background area will be suppressed, so that the network can more accurately find the position of the turnout target during subsequent object detection.

[0027] Through the close cooperation of the network feature extraction and the attention mechanism module, the small target localization strategy can significantly improve the detection accuracy and localization accuracy of small turnout targets in complex track scenarios.

[0028] Image processing steps: Use a preset wheel rolling trajectory detection model to extract the second boundary area of foreign objects on the wheel in the image to be analyzed, and perform a rough classification of the foreign objects in the image to be analyzed to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects. Select a boundary detection strategy according to the rough classification result in the foreign object image to identify the boundary area as the first boundary area; for the image to be analyzed, use the Canny edge detection algorithm to accurately extract the second boundary area; build a semantic analysis model based on a convolutional neural network (CNN) to perform a preliminary identification of the foreign object type in the image to be analyzed. In the model training stage, collect more than 100,000 image samples containing common foreign objects such as plastic bags, paper towels, and metal fragments, and optimize the network parameters by combining transfer learning techniques. The model can extract and analyze the features of the image, and output the probability values of different categories of foreign objects. For example, the probability of determining a certain foreign object as a plastic bag is 85%, and the probability of a paper towel is 15%, realizing a quick rough classification; according to the rough classification result, select the corresponding boundary detection strategy to identify the boundary area of the foreign object image as the first boundary area, and select the corresponding boundary recognition strategy according to the foreign object type and attributes, which can effectively improve the accuracy and efficiency of boundary recognition.

[0029] Flexible foreign objects include paper towels, plastic bags, and fabrics. When the rough classification result is a flexible foreign object, a probability distribution vector representing the probability values of each specific category in the flexible foreign object is output. There are foreign object type labels, corresponding priority boundary recognition strategies, and standard sample features for each category set in the database. According to the distribution vector of the probability values of each specific category in the rough classification result, query the database to obtain the corresponding boundary detection strategy. When detecting foreign objects on the track turnout, through the semantic analysis strategy in the image processing steps, use a convolutional neural network (CNN) model to analyze the image to be analyzed, and obtain a rough classification result of the foreign object, which is divided into two categories: hard objects and flexible foreign objects. When the foreign object is determined to be a flexible foreign object, the model will output a probability distribution vector representing the probability values of each specific category. For example, assuming that the flexible foreign object includes three categories: paper towel, plastic bag, and fabric, then the probability distribution vector may be expressed as [paper towel probability value, plastic bag probability value, fabric probability value], such as [0.3, 0.6, 0.1]. This vector reflects the model's judgment of the possibility of the current foreign object belonging to different flexible foreign object categories. The generation of the probability distribution vector is based on the model's learning of a large number of labeled images during the training process. During training, the model continuously adjusts its own parameters to minimize the error between the prediction result and the true label. In actual applications, when a new image to be analyzed is input, each layer of the neural network of the model extracts and analyzes the features of the image such as color, shape, and texture, and finally outputs the probability values of each category through the fully connected layer, and combines them to form a probability distribution vector.

[0030] The database stores the foreign object type labels of each specific category, as well as the corresponding priority boundary recognition strategies and standard sample features for each label. For each specific category of flexible foreign objects, its priority boundary recognition strategy is obtained through a large number of experiments and actual case summaries. For example, the boundary recognition strategy for paper towels includes using a texture analysis method based on gray-level co-occurrence matrix (GLCM) to detect low-contrast and irregular edge regions; the boundary recognition strategy for fabrics may focus on analyzing their unique weaving textures and wrinkle patterns, and use a feature extraction method based on local binary pattern (LBP) combined with morphological operations to identify the boundary.

[0031] The standard sample features are also an important part of the database. It contains the image feature data of different flexible foreign objects under various lighting conditions and shooting angles, such as the fiber texture feature parameters of paper towels, the grid pattern feature vectors of plastic bags, and the weaving structure feature descriptions of fabrics. These standard sample features are used to compare with the actual foreign object features extracted in the subsequent detailed feature extraction steps to assist in determining the exact type of the foreign object.

[0032] When the probability value of a certain category in the probability distribution vector is significantly greater than the probability values of other categories, that is, when the following conditions are met:

[0033] where, is the maximum probability value in the probability distribution vector, is the probability value of other categories, is a preset dominance coefficient (for example = 1.5, obtained through machine learning optimization). At this time, the category with the highest probability value is determined to be the category to which the foreign object belongs, and using this category as an index, the corresponding priority boundary recognition strategy is queried in the database. For example, if the probability distribution vector is [0.1, 0.8, 0.1], the probability value of the plastic bag, 0.8, is significantly greater than the probability values of the paper towel and the fabric, meeting the above conditions, then the boundary recognition strategy corresponding to the plastic bag is queried in the database. The boundary recognition strategy for plastic bags in the database may include preferentially detecting high-reflectance and smooth edge regions, and using a method combining threshold segmentation and morphological operations to highlight the boundary contour of the plastic bag in the image.

[0034] When the probability values of multiple specific categories all reach the threshold, the first boundary regions corresponding to foreign objects of each category are obtained according to the recognition strategies of each category respectively. Then, through the image comparison strategy, the boundaries of foreign objects in the foreign object image are analyzed. Based on the image similarity, the first boundary regions are determined, and the probability values are adjusted according to the comparison results. When there are multiple probability values of specific categories that all reach the threshold (for example, the probability distribution vector is [0.35, 0.3, 0.25], and the probability values of paper towels, plastic bags, and fabrics all reach the preset threshold of 0.3), the system will process the image to be analyzed respectively according to the boundary recognition strategy corresponding to each category to obtain the foreign object boundaries.

[0035] Based on the similarity calculation results of multiple specific categories, the region with the highest similarity and conforming to the overall foreign object shape logic is selected and determined as the first boundary region. For example, if the region corresponding to the paper towel category has the highest similarity in the foreign object image, and this region can be reasonably connected with the boundary regions determined by other categories in terms of spatial position and shape, then this region is determined as the final first boundary region. After determining the first boundary region, the features of the extracted first boundary region (such as texture features, shape features) are compared with the features of each standard sample in the database, and combined with the previous rough classification probability values, the probability values of each category are adjusted and optimized.

[0036] The boundary detection strategy includes: Targeted image processing steps. According to the rough classification results, when the foreign object is a paper towel or a fabric, enhance the image contrast and retain the edges, and use multi-scale edge detection combined with morphological operations to strengthen the edges; when the foreign object is a plastic bag, suppress the reflection through the polarization strategy, and use a deep learning-based deblurring model and difference repair the boundary; Paper towels and fabric foreign objects usually have the characteristics of low contrast, irregular shapes, and rich textures, and their boundaries are easily confused with the track background. For such foreign objects, the system adopts the following processing flow: Image enhancement preprocessing. Through the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm, the image contrast is increased by 20% - 30% to enhance the gray-scale difference between the paper towel / fabric and the background. For example, when a white paper towel on the track is detected, this operation can make the paper towel area more prominent in the image. The bilateral filtering algorithm is used to smooth the image, removing noise while retaining edge information. This algorithm adjusts the pixel values by calculating the weights in the spatial domain and the value domain, effectively retaining the detailed features of the paper towel edges.

[0037] Canny operator and multi-scale fusion. Use the Canny operator for edge detection to obtain the initial edge map. Then, combined with the multi-scale analysis method, detect edges at different scales and fuse the results to capture multi-level edge information from fine textures to the overall contour. For example, at a small scale, the wrinkled edges of the paper towel can be detected, and at a large scale, the overall contour of the paper towel can be grasped.

[0038] Optimize the edges through dilation and erosion operations. First, use the dilation operation to connect the broken edges, and then remove the redundant edge noise through the erosion operation to finally obtain a continuous and clear tissue / fabric boundary.

[0039] Foreign objects such as plastic bags are transparent, reflective, and easily deformable. Detecting their boundaries faces problems such as reflective interference and edge blurring caused by transparency. Suppress the reflection through the polarization strategy. The following means can be used for processing: The polarization strategy includes: reflection suppression and deblurring. Utilize the polarization characteristics of the plastic bag reflection. By adjusting the angle of the polarization filter, effectively reduce the surface reflection and enhance the visibility of the texture and edges of the plastic bag itself. For example, when there is strong reflection on the surface of the plastic bag, polarization filtering can reveal the details in the reflective area.

[0040] Enhancement of the Retinex algorithm. Apply the single-scale Retinex (SSR) algorithm to enhance the brightness and color correction of the image, improve the contrast between the transparent area of the plastic bag and the background, and highlight the potential boundary information.

[0041] Deep learning deblurring and boundary repair. Adopt a deblurring model based on the U-Net architecture to restore the blurred boundaries caused by transparency and reflection. By learning a large number of blurred-clear image pairs, this model can predict and restore the true edge shape of the plastic bag. For partially invisible boundaries caused by transparency, use the adjacent frame images or the surrounding environment information to perform difference calculations to infer and repair the missing boundary parts. For example, when a part of the plastic bag merges with the background, analyze the texture and color characteristics of the adjacent area to complete this part of the boundary.

[0042] Model training and inference steps, dynamically adjust the loss weights by combining foreign object type specificity and fuzzy IoU loss. The foreign object type-specific loss includes texture constraint loss for tissues and fabrics, transparent area compensation loss and specular area suppression loss for plastic bags. The fuzzy IoU loss is calculated based on fuzzy set theory to calculate the fuzzy intersection and construct it. To better learn the boundary features of different types of foreign objects, combine the foreign object type-specific loss and the fuzzy IoU loss, and dynamically adjust the loss weights; Tissue / Fabric Texture Constraint Loss: For the rich texture characteristics of tissues and fabrics, design a texture constraint loss function. This function extracts the local binary pattern (LBP) features of the image, calculates the difference between the predicted boundary and the true boundary in the texture space, and prompts the model to pay more attention to texture details, improving the detection accuracy of the boundaries of soft and porous materials. Plastic Bag Transparent Area Compensation Loss: Considering the problem that the transparent areas of plastic bags are difficult to detect, introduce a transparent area compensation loss. This loss function applies weighted penalties to the predicted boundaries of the transparent areas of the model, encouraging the model to more accurately identify the edges of these areas. Plastic Bag Specular Area Suppression Loss: To reduce the interference of specular reflections on boundary detection, design a specular area suppression loss. By analyzing the brightness and color features of the image, identify the specular areas, and apply additional penalties to the boundary prediction errors in these areas, guiding the model to ignore specular interference and focus on the true boundaries. Based on fuzzy set theory, regard the boundary prediction as a fuzzy event, and calculate the fuzzy intersection and union of the predicted boundary and the true boundary. The traditional IoU loss is sensitive to small offsets of the boundary, while the fuzzy IoU loss can better handle the uncertainty of the boundary by introducing the concept of fuzziness, improving the model's tolerance and robustness to inaccurate boundaries.

[0043] Construct a multi-task learning network to optimize both the foreign object classification and boundary detection tasks simultaneously. By sharing the feature extraction layer, the model can learn semantic information related to the boundary, improving the accuracy of boundary detection.

[0044] Dynamically adjust the weights of various loss functions according to the detection difficulty of different types of foreign objects during training. For example, when the boundary detection error of plastic bags is large, automatically increase the weights of the loss functions related to plastic bags, prompting the model to pay more attention to the boundary features of such foreign objects.

[0045] The model generates an initial boundary prediction result based on the input image, and then combines the enhanced edge information obtained in the targeted image processing steps to optimize and refine the predicted boundary. Remove redundant boundary predictions through the non-maximum suppression (NMS) algorithm, and then evaluate the confidence of the final boundary detection result according to the result calculated by the fuzzy IoU loss to ensure that the output boundary is accurate and reliable.

[0046] Through the above targeted image processing and model training and inference strategies, the track turnout foreign object boundary detection system can effectively cope with the challenges of different types of foreign objects and achieve high-precision boundary detection of common foreign objects such as paper towels, fabrics, and plastic bags.

[0047] In the detailed feature extraction step, corresponding texture features are respectively extracted in the first boundary region and the second boundary region through a deep learning network, and the extracted texture features are mapped one by one to obtain a mapping relationship group. Then, the feature change rate is obtained through the analysis of the mapping relationship group. The feature change rate includes the texture contrast ratio, the color contrast ratio, and the shape contrast ratio. Then, the specific foreign object type is obtained through the multi-feature analysis strategy based on the texture contrast ratio, the color contrast ratio, and the shape contrast ratio. Using a convolutional neural network (such as ResNet, DenseNet), the images of the first boundary region and the second boundary region are respectively analyzed and the texture features are respectively extracted. The pixel-level similarity between the first boundary region and the second boundary region is calculated through an assistant extraction mechanism to generate a mapping weight matrix. For each feature point in the first boundary region, the corresponding point with the most similar texture is found in the second boundary region to form a set of "feature point pairs". The weight matrix reflects the correlation strength of the features in different regions and suppresses the interference of background noise. The texture features include foreign object texture features, foreign object color features, and foreign object deformation features. The network will learn and capture the fine textures on the surface of foreign objects, such as the fiber texture of paper towels, the folds of plastic bags, and the scratches on the metal surface. At the same time, by comparing the texture features of the two groups of images, the texture changes of foreign objects from different perspectives are analyzed, such as whether there is texture blur caused by reflection, whether there are new textures generated by rolling, etc. A network structure combined with spatial pyramid pooling (SPP) is used to analyze the shape changes of foreign objects at multiple scales. The system not only calculates the overall geometric indexes such as the aspect ratio and roundness of foreign objects, but also focuses on local deformations, such as the bending angle of branches and the stretching degree of plastic bags. By comparing the shape contours of the two boundary regions, it is judged whether the foreign object has dynamic changes such as extrusion and distortion when the train passes by. The texture features are respectively analyzed through the mapping relationship group to obtain the feature change rate, and the feature change rate includes the image feature change rate.

[0048] The multi-feature analysis strategy includes combining the feature change rate with the probability weight of coarse classification and using a weighted voting mechanism to output the foreign object type. When there is a conflict between the feature change rate and the coarse classification result, the weight is dynamically adjusted according to the credibility of different features.

[0049] After the image processing step is completed, the system extracts the probability values of each specific category in the coarse classification result from the result output by the semantic analysis strategy. For example, when the coarse classification determines that the foreign object is a flexible foreign object, a distribution vector containing the probabilities of categories such as paper towels, plastic bags, and fabrics will be output. For example, the probability vectors of paper towels, plastic bags, and fabrics are [0.3, 0.5, 0.2].

[0050] From the first boundary region of the foreign object image, texture features are extracted using methods such as Local Binary Pattern (LBP) and Gray-Level Co-Occurrence Matrix (GLCM). Taking LBP as an example, by comparing the gray-scale relationship between the central pixel and neighboring pixels, a binary code is generated to describe the texture details (such as the fiber structure of a tissue paper and the grid pattern of a plastic bag). The extracted texture feature vector is compared with the standard sample texture features of the corresponding foreign object type in the database for similarity calculation. The commonly used cosine similarity formula is:

[0051] The calculated is the texture contrast ratio. The closer its value is to 1, the more similar the texture features are to the standard sample.

[0052] The first boundary region of the foreign object image is compared with the second boundary region of the image to be analyzed. The change rates of geometric parameters such as length, width, and area are calculated, and at the same time, shape change features such as the degree of fold and edge distortion of the indentation are extracted. For example, the area change rate is calculated

[0053] is the area after pressing, is the area before pressing. The similarity between these deformation features and the contour of the corresponding standard sample in the database is calculated. Cosine similarity or Euclidean distance measurement can also be used to obtain the deformation feature value. Combining the contour features (such as perimeter and concavity) of the first boundary region further enhances the accuracy of deformation analysis.

[0054] The first boundary region and the second boundary region are respectively subjected to histogram equalization to improve the stability of image features, and the mean, variance, and skewness of the RGB three channels are calculated to describe the concentration region and dispersion degree of color distribution. Combining the above feature change rates, a color contrast ratio is comprehensively obtained. The larger the value of the color contrast ratio, the more obvious the color difference between the first boundary region and the second boundary region, which can reduce the color change and hue shift caused by foreign object reflection.

[0055] Vote weights are assigned to each specific type according to the probability values of the coarse classification. The higher the probability value of a category, the greater its weight, to reflect the priority of this category in the preliminary judgment. Weights are assigned to the texture feature contrast ratio, color contrast ratio, and shape contrast ratio to highlight the importance of different features in classification. For each possible foreign object type, its vote score is calculated, and the vote scores calculated for all categories are compared. The category with the highest score is selected as the final foreign object type output. And when there are obvious conflicts among the texture contrast ratio, color contrast ratio, shape contrast ratio, or the features of the coarse classification result, according to the different dependencies of different foreign object features on the feature change rates, the feature change rates are adjusted. When the surface texture is obviously a tissue, the weight ratio of the texture feature is increased. For a plastic bag that is prone to reflection, the weight ratio of the shape feature is increased to reduce the deviation effects caused by the reflection interference on the texture feature and the shape contrast feature.

[0056] It also includes a position recognition step. The position of the foreign object is judged according to the contour lines of the first boundary region and the contour lines of the track turnout, and the deformation trend of the foreign object is analyzed according to the contour lines of the second boundary region and the contour lines of the first boundary region. The position of the deformed foreign object is analyzed, and a danger warning is generated according to the foreign object type. After the foreign object boundary recognition in the image processing step is completed, the Suzuki contour tracking algorithm is used to process the first boundary region determined in the foreign object image. This algorithm starts from the edge pixel points of the image and, through the method of eight-neighborhood search, tracks the boundary pixels in the clockwise or counterclockwise direction to gradually construct a complete contour line. During the tracking process, the algorithm records the coordinates of each pixel point, and finally forms a pixel coordinate sequence describing the shape and range of the foreign object before being pressed. To improve the accuracy of the contour line, a morphological thinning operation can be performed on the boundary recognition result to remove the burrs and redundant pixels on the edge, making the contour smoother and more accurate. The system pre-stores a high-precision standard contour model of the track turnout, which is constructed based on the track design drawings and actual survey data and contains the accurate geometric information of each component (switch rail, stock rail, frog, etc.) of the turnout. During real-time detection, first, the image registration technology based on feature point matching is used to extract SIFT or SURF feature points from the collected track turnout image, match them with the feature points in the standard model, calculate the rotation, translation, and scaling parameters of the image, and align the real-time image with the standard model.

[0057] Subsequently, the Canny edge detection algorithm is used to extract the actual contour line of the track turnout. Due to interference factors such as oil stains, wear, and uneven illumination on the track surface, before performing Canny detection, the image is first processed by adaptive histogram equalization to enhance the contrast of the image. At the same time, according to the characteristics of the track image, the high and low thresholds of the Canny algorithm are dynamically adjusted (for example, by calculating the statistical distribution of the image gradient, the thresholds are automatically set to 1.5 times and 3 times the gradient mean) to ensure that both the turnout contour can be completely extracted and noise interference can be effectively suppressed. Finally, morphological closing operation is used to connect the broken parts on the contour line, and opening operation is used to remove small noise regions to obtain the accurate contour line of the track turnout. The track turnout area is divided into multiple sub-regions such as the switch rail area, frog area, guard rail area, and the area near the switch machine according to functional and structural characteristics, and the corresponding coordinate range and geometric model are established for each sub-region. Calculate the geometric center (centroid) coordinates of the first boundary region contour line, and the calculation formula is:

[0058] where $x_i$ and $y_i$ are the coordinates of the $i$-th point on the contour line, and $n$ is the total number of contour points. By determining which sub-region the centroid coordinates fall into, the approximate location of the foreign object can be quickly determined. Calculate the overlapping area between the contour line of the first boundary region and the contour line of the track turnout to quantify the impact of the foreign object on track operation. Using the pixel counting method, convert the regions where the two contour lines are located into binary images, obtain the binary image of the overlapping part through logical AND operation, count the number of pixels in the overlapping region, and then convert it into the actual area according to the actual resolution of the image. If the overlapping area exceeds the preset threshold (such as 30%), it indicates that the foreign object poses a greater threat to the normal operation of the track, and the risk level assessment needs to be improved. Utilize the principles of affine transformation and perspective transformation in image processing to align the contour line of the second boundary region (after pressing) with the contour line of the first boundary region (before pressing) to accurately compare the shape and position changes between the two. During the specific operation, first calculate the central position and the main axis direction of the two contour lines through contour moments, and then construct an affine transformation matrix based on the offset of the central position and the rotation angle of the main axis direction to transform the contour line of the second boundary region to the same coordinate system as the contour line of the first boundary region. Calculate the displacement vector between the two contour lines, including the translation distance and the rotation angle. The translation distance can be obtained by calculating the Euclidean distance between the center of the transformed second boundary region contour line and the center of the first boundary region contour line; the rotation angle is determined according to the change angle of the main axis direction. At the same time, analyze the extension direction of the indentation, for example, by calculating the tangent direction of the indentation contour line, to judge the sliding or rolling trend of the foreign object under the rolling of the train. If the contour line of the second boundary region translates towards the track center direction and rotates relative to the contour line of the first boundary region, it indicates that the foreign object moves towards the key area of the track after being pressed, which may affect the normal conversion of the turnout or the wheel-rail contact of the train. At this time, the risk level needs to be further increased; the foreign object type is a flexible foreign object (such as a plastic bag, fabric, etc.), and it completely covers the key parts of the track (such as the close contact area between the switch rail and the stock rail, the throat area of the frog). Although the flexible foreign object is relatively soft in texture, under the action of the train airflow and wheel rolling, it may be wrapped around the turnout components, affecting the normal conversion of the turnout and causing the train running direction to get out of control. After the system triggers a high-risk warning, the emergency braking operation is also executed, and the ground control center is notified for handling; if the flexible foreign object partially covers the track but does not completely block it and is located in a non-critical area (such as the guard rail area, the track edge), a medium-risk warning is generated. At this time, the system sends a deceleration operation instruction to the train to reduce the train speed to a safe speed (such as 20 km / h) to reduce the impact of the foreign object on train operation. At the same time, arrange inspection personnel to carry tools to go to the scene to clean the foreign object as soon as possible through the dispatching system, and strengthen the monitoring of this area during the subsequent operation process; for small foreign objects with a small volume and located in a remote location, such as small pieces of paper and small stones at the track edge, a low-risk warning is issued. The system records the location information, type, and image data of the foreign object and incorporates it into the daily inspection plan.During the next track inspection, the inspection personnel will clean it and conduct a key inspection of this area to ensure that foreign objects do not pose a potential threat to the operation of the track.

[0059] As Figure 4 shown, there is also a verification step. There is also a verification module on the train. The verification module is arranged behind the second camera group. The verification module includes a main camera, an auxiliary camera, and a lidar. According to the type of foreign object, the lighting intensity and camera parameters of the auxiliary camera are adjusted, and the image data collected by the main camera and the auxiliary camera and the laser electrical signals collected by the lidar are combined. The data is verified based on multi-modal data fusion processing, and based on the verification result, an abnormal analysis result is obtained and compared with the foreign object type to determine whether they are consistent. When the target detection model preliminarily determines the foreign object type, the verification module will immediately automatically adjust the working parameters of the auxiliary camera according to the characteristics of the foreign object. The auxiliary camera is equipped with a time-sharing stroboscopic function. Using a group of light sources of different colors, by precisely controlling the stroboscopic of each light source at different time points, the image information under different color light sources is obtained, and according to the object attributes, the image photos under different lighting effects are obtained.

[0060] For foreign objects with strong reflectivity: If the detected foreign object is a metal object, the surface of this type of object is smooth and prone to reflection, resulting in over-bright images and loss of details. At this time, the auxiliary camera will reduce the lighting intensity to avoid the problem of reflection being exacerbated by strong light irradiation, ensuring that details such as the texture and contour of the metal foreign object can be clearly presented in the image.

[0061] For foreign objects with low color contrast: For foreign objects with light colors and similar colors to the background, such as white paper towels, it is very difficult to distinguish them against the track background. The auxiliary camera will select the image photo under warm colors, enhance the color difference between the foreign object and the background by selecting different lighting colors; at the same time, it will also optimize parameters such as the exposure time and contrast of the camera to make the white paper towel more prominent in the image for subsequent analysis.

[0062] For other types of foreign objects: For foreign objects of different materials and shapes, such as plastic bags, branches, etc., the auxiliary camera will also adjust the parameters accordingly. For example, for a thin and transparent plastic bag, the focusing parameter will be adjusted to ensure that its edge is clear; for an irregularly shaped foreign object like a branch, the shooting angle and field of view will be adjusted to completely capture its shape.

[0063] Analyze the dual-camera images using deep learning algorithms (such as ResNet+FPN) to extract more than 20 visual features such as edge contours, texture features (LBP algorithm), and color histograms; perform noise reduction processing (voxel filtering) on the lidar point cloud data, and calculate geometric features such as point cloud density, surface curvature, and normal vectors to construct a three-dimensional space description model. Project the point cloud data onto the image plane to generate a depth image with spatial information, realizing the preliminary association of visual and spatial data; classify and predict the image features and point cloud features respectively, and then integrate the results through a weighted voting algorithm to finally output a unified foreign object description vector.

[0064] Match the fused foreign object description vector with the standard feature template in the system database: Match successful: If the similarity > 90% and is consistent with the preliminary judgment of the target detection model, the verification passes, and the system records the foreign object information and generates a regular alarm; Match failed: Trigger exception handling when the following situations occur: Feature similarity < 80%; or there are contradictions in multi-modal data (such as visual judgment being plastic while lidar shows metal reflection characteristics), conflicting with the preliminary judgment of the foreign object type.

[0065] At this time, the system immediately starts the secondary detection program, calls historical data for comparative analysis; if it still cannot be confirmed, it sends a high-priority alarm to the train dispatching center, attaches detailed multi-modal data, and prompts manual intervention for verification.

[0066] As Figure 3 shown, a track switch foreign object detection and category analysis system includes: A switch foreign object recognition module, with a first camera group set at the front end of the train, uses the first camera group to shoot track videos, performs frame extraction detection, then identifies the switch position through a preset target detection model, calibrates the foreign objects in the switch through a preset foreign object detection model, and intercepts the corresponding images in the video as foreign object images; An image acquisition module, uses a second camera group set behind the train wheels to shoot the images of the foreign objects calibrated in the switch as the images to be analyzed; An image processing module, extracts the second boundary area where the foreign object is run over by the wheels in the image to be analyzed through a preset wheel rolling trajectory detection model, and roughly classifies the foreign objects in the image to be analyzed to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects, and selects a boundary detection strategy according to the rough classification result in the foreign object image to identify the boundary area as the first boundary area; The detail feature extraction module extracts corresponding texture features in the first boundary region and the second boundary region respectively through a deep learning network, maps the extracted texture features one by one to obtain a mapping relationship group, and then analyzes the feature change rate through the mapping relationship group. The feature change rate includes a texture contrast rate, a color contrast rate, and a shape contrast rate. Then, according to the texture contrast rate, the color contrast rate, and the shape contrast rate, a specific foreign object type is analyzed through a multi-feature analysis strategy.

[0067] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting foreign objects and analyzing their types in a railway turnout, characterized in that, Including: The turnout foreign object recognition step: A first camera group is set at the front end of the train. The first camera group is used to shoot track videos, perform frame extraction detection, then identify the turnout position through a preset target detection model, calibrate the foreign objects in the turnout through a preset foreign object detection model, and intercept the corresponding image in the video as the foreign object image; The image acquisition step: A second camera group set behind the train wheels is used to shoot the image of the foreign object calibrated in the turnout as the image to be analyzed; The image processing step: A second boundary area of the foreign object being run over by the wheel is extracted from the image to be analyzed through a preset wheel rolling trajectory detection model, and the foreign object in the image to be analyzed is roughly classified to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects. According to the rough classification result in the foreign object image, a boundary detection strategy is selected to identify the boundary area as the first boundary area; The detailed feature extraction step: Corresponding texture features are respectively extracted from the first boundary area and the second boundary area through a deep learning network, and the extracted texture features are mapped one by one to obtain a mapping relationship group. Then, the feature change rate is analyzed through the mapping relationship group. The feature change rate includes a texture contrast rate, a color contrast rate, and a shape contrast rate. Then, according to the texture contrast rate, the color contrast rate, and the shape contrast rate, a specific foreign object type is analyzed through a multi-feature analysis strategy.

2. The method for detecting foreign objects and analyzing types of a track turnout according to claim 1, wherein The flexible foreign objects include paper towels, plastic bags, and fabrics. When the rough classification result is a flexible foreign object, a probability distribution vector representing the probability values of each specific category in the flexible foreign object is output. A foreign object type label, a corresponding priority boundary recognition strategy for each category, and standard sample features are set in the database. According to the distribution vector situation of the probability values of each specific category in the rough classification result, the database is queried to obtain the corresponding boundary detection strategy.

3. The method for detecting foreign objects and analyzing their types in a track turnout according to claim 2, wherein, When the probability values of multiple specific categories all reach the threshold, the first boundary area corresponding to each category of foreign object is obtained according to the recognition strategy of each category respectively, and the foreign object boundary in the foreign object image is analyzed through an image comparison strategy. According to the image similarity, the first boundary area is determined, and the probability value is adjusted according to the comparison result.

4. The method for detecting foreign objects and analyzing types on a track turnout according to claim 2, wherein, The boundary detection strategy includes: The targeted image processing step: According to the rough classification result, when the foreign object is a paper towel or a fabric, the image contrast is enhanced and the edges are retained, and multi-scale edge detection is combined with morphological operations to strengthen the edges; when the foreign object is a plastic bag, the polarization strategy is used to suppress the reflection, and a deblurring model based on deep learning and a difference repair boundary are used; The model training and inference step: The loss weight is dynamically adjusted by combining the foreign object type specificity and the fuzzy IoU loss. The foreign object type specificity loss includes a texture constraint loss for paper towels and fabrics and a transparent area compensation loss and a reflection area suppression loss for plastic bags. The fuzzy IoU loss is calculated based on the fuzzy set theory to calculate the fuzzy intersection and construct.

5. A method for detecting foreign objects and analyzing their types in a track turnout, as claimed in claim 1, wherein It further includes a position recognition step, which determines the location of the foreign object according to the contour lines of the first boundary region and the track turnout, analyzes the deformation trend of the foreign object based on the contour lines of the second boundary region and the first boundary region, analyzes the location of the deformed foreign object, and generates a danger warning according to the type of the foreign object.

6. The foreign object detection and type analysis method for a track turnout according to claim 1, characterized in that, The multi-feature analysis strategy includes combining the feature change rate with the probability weight of rough classification, and using a weighted voting mechanism to output the type of foreign object. When there is a conflict between the feature change rate and the rough classification result, the weight is dynamically adjusted according to the credibility of different features.

7. A method for detecting foreign objects and analyzing their types in a track turnout, as claimed in claim 1, wherein It further includes a verification step. A verification module is also provided on the train. The verification module is arranged behind the second camera group. The verification module includes a main camera, an auxiliary camera and a lidar. According to the type of the foreign object, the lighting intensity and camera parameters of the auxiliary camera are adjusted, and the image data collected by the main camera and the auxiliary camera and the laser electric signal collected by the lidar are combined. The data is verified based on multi-modal data fusion processing, and an abnormal analysis result is obtained based on the verification result and compared with the type of the foreign object to determine whether they are consistent.

8. A method for detecting foreign objects and analyzing their types in a railway turnout according to claim 1, characterized in that, A small target localization strategy is set in the target detection model, including using a network architecture for feature extraction, and embedding an attention mechanism module in the network. The attention mechanism module includes a dynamic channel weight module, which assigns high weights to the key feature channels of the turnout data set through the SE-Net structure and performs cross-scale channel interaction to screen the key features of the turnout target. And through a spatial attention module, the global position of the turnout is captured through global pooling to generate an attention map to locate the target area.

9. The method for detecting foreign objects and analyzing types on a track turnout according to claim 1, wherein The second camera group is a high-definition camera. When the first camera group detects the turnout, the second camera is adjusted so that the second camera captures an image of the area where the turnout is located.

10. A foreign object detection and classification analysis system for a track turnout, characterized in that, It includes: A turnout foreign object recognition module. A first camera group is provided at the front end of the train. The first camera group is used to capture the track video, perform frame extraction detection, then identify the turnout position through a preset target detection model, calibrate the foreign object in the turnout through a preset foreign object detection model, and intercept the corresponding image in the video as the foreign object image; An image acquisition module, which captures the image of the foreign object calibrated in the turnout through the second camera group arranged behind the train wheels as the image to be analyzed; An image processing module, which extracts the second boundary region of the foreign object on the wheel being rolled by a preset wheel rolling trajectory detection model in the image to be analyzed, and performs rough classification on the foreign object in the image to be analyzed to obtain a rough classification result. The rough classification result includes hard objects and flexible foreign objects. According to the rough classification result in the foreign object image, a boundary detection strategy is selected to identify the boundary region as the first boundary region; The detailed feature extraction module extracts corresponding texture features in the first boundary region and the second boundary region respectively through a deep learning network, maps the extracted texture features one by one to obtain a mapping relationship group, and then analyzes the feature change rate through the mapping relationship group. The feature change rate includes a texture contrast rate, a color contrast rate, and a shape contrast rate. Then, according to the texture contrast rate, the color contrast rate, and the shape contrast rate, a specific foreign object type is analyzed through a multi-feature analysis strategy.

Images