Subway line foreign matter invasion identification method

Abstract

The invention discloses a subway line foreign matter invasion identification method comprising the following steps: arranging trackside monitoring, positioning, calculating and communication devices along a subway track line, and arranging an image detection device and a radar detection device at the head of a subway vehicle for data acquisition; then performing identification detection on a real-time view image in the collected data to obtain a view image feature detection result; identifying and monitoring the real-time radar data to obtain a current radar detection object position; according to the data synchronization frequency, performing target position fusion on the current radar detection object position after coordinate conversion and the view image feature detection result to obtain a vehicle-mounted intrusion detection result; and uploading a vehicle-mounted intrusion detection result by using a subway vehicle-mounted communication device, and carrying out mapping alarm on a subway track line simulation map in combination with a trackside monitoring foreign matter intrusion result.

Description

Technical Field

[0001] The present invention belongs to the technical field of subway track safety. More specifically, it relates to a method for identifying foreign object intrusion in subway lines. Background Art

[0002] Intrusion specifically refers to the situation where an object exceeds the allowable safety contour size range within the safety limit of rail transit operation, thus interfering with the normal operation of trains on the line in this section. In recent years, urban rail transit accidents caused by object intrusion have been frequent.

[0003] Urban rail transit lines are usually laid in tunnels. Affected by factors such as route selection restrictions, there are many small-radius curves and large slopes, and the slope changes greatly, with a complex line type, which causes many difficulties in the development of train autonomous obstacle detection systems. Currently, most urban rail transit lines detect obstacle intrusion through millimeter-wave radars, lidars, visual sensors, acoustic sensors, etc. Although it can reduce the occurrence of missed detections of obstacles, its effective detection distance is short, and there are blind spots, and there are still problems with insufficient accuracy and efficiency in obstacle recognition and positioning. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for identifying foreign object intrusion in subway lines, which solves the problem that the existing foreign object intrusion detection in urban rail transit is not comprehensive and accurate enough.

[0005] To achieve the above-mentioned invention purpose, a method for identifying foreign object intrusion in subway lines of the present invention is characterized by including the following steps:

[0006] (1) Set up trackside monitoring, positioning, calculation, and communication devices along the subway track line respectively, used to identify and upload the results of trackside monitoring of foreign object intrusion, and construct a simulated map of the subway track line;

[0007] (2) Set up an image detection device and a radar detection device at the head of the subway car, used to capture real-time vision images in the subway traveling direction, and detect real-time radar data in the subway traveling direction;

[0008] (3) Use the vision image foreign object detection neural network in the subway vehicle-mounted calculation device to identify and detect the real-time vision image, and obtain the vision image feature detection result;

[0009] (4) According to the Kalman filtering algorithm, use the subway vehicle-mounted calculation device to identify and monitor the real-time radar data, and obtain the current position of the radar-detected object;

[0010] (5) According to the data synchronization frequency, the on-vehicle computing device of the subway uses the current radar-detected object position and the detection result of the vision image feature to perform target position fusion to obtain the on-vehicle intrusion detection result;

[0011] (6) The on-vehicle computing device of the subway combines the on-vehicle intrusion detection result and the trackside monitoring foreign object intrusion result to perform mapping and warning on the subway track line simulation map.

[0012] The invention purpose of the present invention is achieved as follows:

[0013] A method for identifying foreign object intrusion on a subway line in the present invention sets trackside monitoring, positioning, computing and communication devices along the subway track line, and an image detection device and a radar detection device at the head of the subway car for data collection; then, the real-time vision image in the collected data is identified and detected to obtain the detection result of the vision image feature; the real-time radar data is identified and monitored to obtain the current radar-detected object position; according to the data synchronization frequency, the current radar-detected object position after coordinate conversion and the detection result of the vision image feature are subjected to target position fusion to obtain the on-vehicle intrusion detection result; the on-vehicle intrusion detection result is uploaded by using the on-vehicle communication device of the subway, and combined with the trackside monitoring foreign object intrusion result, mapping and warning are performed on the subway track line simulation map.

[0014] Meanwhile, a method for identifying foreign object intrusion on a subway line in the present invention also has the following beneficial effects:

[0015] (1) By setting trackside monitoring, positioning, computing and communication devices along the subway track line, and an image detection device and a radar detection device at the head of the subway car in the present invention, all-round and multi-dimensional monitoring of foreign objects in the subway traveling direction is realized;

[0016] (2) By combining trackside monitoring and on-vehicle monitoring in the present invention, the accuracy and reliability of foreign object detection are improved, and safety accidents caused by foreign object intrusion are effectively reduced;

[0017] (3) By using the foreign object detection neural network of the vision image to identify and detect the real-time vision image, the Kalman filter algorithm to process the radar data, and based on the data synchronization frequency, performing target position fusion on the radar-detected object position after coordinate conversion and the detection result of the vision image feature in the present invention, the accuracy and precision of foreign object detection are improved;

[0018] (4) By constructing a subway track line simulation map and mapping warning information on the map in the present invention, visual management of foreign object intrusion is realized, which is convenient for intuitively understanding the situation of foreign object intrusion, quickly making decisions and taking corresponding countermeasures, and effectively improving the emergency response speed and management efficiency. Description of the Drawings

[0019] Figure 1 is a flowchart of a method for identifying foreign object intrusion on a subway line according to the present invention;

[0020] Figure 2 is a block diagram of a monitoring system for foreign object intrusion on a subway line according to the present invention. Specific embodiments

[0021] The following describes the specific embodiments of the present invention with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0022] Embodiment

[0023] Figure 1 is a flowchart of a method for identifying foreign object intrusion on a subway line according to the present invention.

[0024] In this embodiment, as Figure 1 shown, a method for identifying foreign object intrusion on a subway line according to the present invention includes the following steps:

[0025] (1) Set trackside monitoring, positioning, calculation, and communication devices along the subway track line respectively, for identifying and uploading the results of foreign object intrusion monitored trackside, and constructing a simulated map of the subway track line;

[0026] (1.1) Set trackside monitoring devices and trackside calculation devices connected to the trackside monitoring devices along the platforms, curves, and connecting channels of the subway track line at a preset first interval distance. Among them, the trackside monitoring devices are used to capture trackside monitoring images; the trackside calculation devices are used to identify the results of foreign object intrusion in the trackside monitoring images;

[0027] In this embodiment, the trackside monitoring devices adopt long-focus laser high-definition cameras, and the first interval distance can enable the trackside monitoring devices to cover the intrusion-prone areas such as platforms, curves, and connecting channels without dead angles; considering the foreign object intrusion situations of different scales, in order to avoid false detection caused by large objects after magnification and missed detection of intrusion targets caused by small objects, the trackside monitoring images are subjected to enhanced object detection at different scales, so that the trackside calculation devices can quickly and accurately identify the results of foreign object intrusion in the trackside monitoring images. The specific method is as follows:

[0028] (1.1.1) Obtain a number of trackside monitoring images to form a trackside monitoring image dataset;

[0029] (1.1.2) Keep the backbone network and neck network of the YOLOv5 network unchanged, replace the three prediction head sub-networks in the prediction head network with a classification prediction sub-network, a size prediction sub-network, and an offset prediction sub-network to obtain a trackside image intrusion recognition network, and then set the trackside image intrusion recognition network in the trackside computing device;

[0030] In this embodiment, the prediction head network uses three parallel convolutional blocks as the classification prediction sub-network, the size prediction sub-network, and the offset prediction sub-network respectively, and outputs three feature maps with the same resolution. Through the feature map output by the classification prediction sub-network, the category and location information of the intrusion target can be predicted. Among them, different channels represent different categories, and the matrix element value in the feature map matrix is the confidence score that the corresponding point is the center of the intrusion target; the feature map output by the size prediction sub-network has two channels, and the corresponding values are the width and height of the intrusion target; the feature map output by the offset prediction sub-network has two channels, and the corresponding values are the offsets of the center pixel of the intrusion target in the horizontal and vertical directions;

[0031] (1.1.3) According to the multi-scale object loss function, input the trackside monitoring image into the trackside image intrusion recognition network for intrusion target recognition prediction training to obtain a trained trackside image intrusion recognition network;

[0032] In this embodiment, the multi-scale object loss function used by the trackside image intrusion recognition network during training is:

[0033] L = L1 + μ1L2 + μ2L3

[0034]

[0035] where L represents the multi-scale object loss function, L1 represents the classification loss function, μ1 represents the first loss function hyperparameter, L2 represents the offset loss function, μ2 represents the second loss function hyperparameter, L3 represents the size loss function, N represents the total number of intrusion targets, C represents the number of channels, H represents the height of the trackside monitoring image, W represents the width of the trackside monitoring image, c represents the category of the pixel in the image, y represents the pixel ordinate of the image, x represents the pixel abscissa of the image, represents the probability that the pixel belongs to the center point of a certain category, υ cyx represents the true value of the probability that the pixel belongs to the center point of a certain category, α represents the first category index factor, β represents the second category index factor, represents the offset between the center point of the i-th intrusion target in the trackside monitoring image and the center point of the true target box in the trackside monitoring image, θ idenotes the true value of the offset between the center point of the $i$-th intrusion target in the on-rail monitoring image and the center point of the true target box in the on-rail monitoring image, and $\|\cdot\|_1$ represents the calculation of the first norm. denotes the size of the $i$-th intrusion target corresponding to the predicted center point, $\xi$ i denotes the true size of the $i$-th intrusion target.

[0036] (1.1.4) Use the trained on-rail image intrusion recognition network to perform intrusion target recognition on the newly captured on-rail monitoring image to obtain the original-scale intrusion target recognition result. Among them, the original-scale intrusion target recognition result includes the position, size, and category of each intrusion target;

[0037] (1.1.5) Determine whether there is an intrusion target with a size greater than the preset size threshold in the original-scale intrusion target recognition result. If so, regard the newly captured on-rail monitoring image as a large-target on-rail monitoring image, and then go to step (1.1.6); otherwise, jump to step (1.1.7);

[0038] (1.1.6) Set the image scaling factor $\eta$ s as the hyperparameter for image resolution adjustment; according to the image scaling factor perform a shrinking process on the newly captured on-rail monitoring image to obtain a scale-reduced on-rail monitoring image; use the trained on-rail image intrusion recognition network to perform intrusion target recognition on the scale-reduced on-rail monitoring image to obtain a scale-reduced intrusion target recognition result; finally, take the union of the scale-reduced intrusion target recognition result and the original-scale intrusion target recognition result to obtain the final on-rail monitoring foreign object intrusion result;

[0039] (1.1.7) Set the image scaling factor According to the image scaling factor perform a shrinking process and an enlarging process on the newly captured on-rail monitoring image respectively to obtain a scale-reduced on-rail monitoring image and a scale-enlarged on-rail monitoring image; use the trained on-rail image intrusion recognition network to perform intrusion target recognition on the scale-reduced and scale-enlarged on-rail monitoring images respectively to obtain a scale-reduced and scale-enlarged intrusion target recognition result and a scale-enlarged intrusion target recognition result; finally, take the union of the scale-reduced intrusion target recognition result, the scale-enlarged intrusion target recognition result, and the original-scale intrusion target recognition result to obtain the final on-rail monitoring foreign object intrusion result;

[0040] In this embodiment, when performing target scale reduction recognition on the on-rail monitoring image, the resolution of the scale-reduced on-rail monitoring image is $1 - \eta$ times the resolution of the newly captured on-rail monitoring image. stimes; when performing target scale magnification recognition on the trackside monitoring image, the resolution of the magnified trackside monitoring image is 1 + η times that of the newly captured trackside monitoring image s times; By processing the intrusion situations of foreign objects at different scales, false detections caused by large objects after magnification and missed detections of intrusion targets caused by small objects are avoided;

[0041] (1.2) Set trackside positioning devices and trackside communication devices along the subway track line in sequence and simultaneously according to a preset second interval distance. Among them, the trackside positioning devices and trackside computing devices at platforms, curves, and connection passages are all connected to the trackside communication device, and are used to transmit the trackside monitoring foreign object intrusion results combined with the line positioning information to the subway line security center;

[0042] In this embodiment, the second interval distance is less than or equal to the maximum communication distance of the trackside communication device. The trackside positioning device is a trackside positioning identification card, and the trackside communication device includes a LET module and a Lora module; the trackside positioning identification card is used to divide the track line sections to accurately provide the positioning information of the railway line sections. Combined with the foreign object intrusion results, the foreign object intrusion position can be quickly located, providing a basis for timely clearing line obstacles and ensuring traffic.

[0043] (1.3) Construct a line simulation map according to the subway track line, and map the actual installation positions of the trackside monitoring device, trackside computing device, trackside positioning device, and trackside communication device into the line simulation map to obtain the subway track line simulation map.

[0044] (2) Set the image detection device and the radar detection device at the front of the subway car, which are used to capture the real-time field of view image in the subway traveling direction and detect the real-time radar data in the subway traveling direction;

[0045] In this embodiment, the image detection device uses a vision sensor, and the radar detection device uses a millimeter-wave radar; the acquisition processes of the real-time field of view image and the real-time radar data are as follows:

[0046] (2.1) Set the image detection device at the center position of the front of the subway car, and use the image detection device to capture the real-time field of view image in the subway car traveling direction;

[0047] (2.2) Set the radar detection device above the image detection device so that the radar detection device is located on the central axis of the track line; then, with the radar detection device as the origin, the subway traveling direction as the longitudinal coordinate axis direction, and the direction perpendicular to the track line as the transverse coordinate axis direction, a rectangular coordinate system is constructed;

[0048] (2.3) Detect the distance, angle, and relative velocity of an object in the advancing direction of the subway train using a radar detection device based on the difference frequency of the Doppler shift until the object is detected.

[0049] In this embodiment, detecting an object means that the radar signal continuously detects a certain object exceeding a preset detection threshold. If the number of times of detecting the object is less than the preset detection threshold, it is considered that the object is lost or the detection fails, and the process of processing the detection data in the next step is not entered. Only when the number of times of detecting the object exceeds the preset detection threshold, the process of processing the detection data in the next step is entered. In this embodiment, the radar detection device uses a triangular wave modulation signal for foreign object detection.

[0050] (2.4) Number the detected objects, and combine the numbers of the objects with the corresponding distances, angles, and relative velocities of the objects to obtain the real-time radar data in the advancing direction of the subway.

[0051] (3) Use the foreign object detection neural network in the subway vehicle computing device to identify and detect the real-time field of view image to obtain the field of view image feature detection result.

[0052] (3.1) Obtain a number of field of view images with or without foreign objects, and perform weighted average graying processing and Gaussian filtering processing on the field of view images in sequence to obtain the preprocessed field of view images.

[0053] In this embodiment, the model of the weighted average graying processing is:

[0054] Gray(a,b)=(ω R R(a,b)+ω G G(a,b)+ω B B(a,b))

[0055] Among them, Gray(a,b) represents the gray value of the pixel at position (a,b) in the field of view image, ω R 、ω G 、ω B respectively represent the weight coefficients of the red, green, and blue channels, R(a,b), G(a,b), and B(a,b) respectively represent the red, green, and blue channel pixel values of the pixel at position (a,b) in the field of view image, a is the abscissa of the field of view image, and b is the ordinate of the field of view image;

[0056] In this embodiment, the model of the Gaussian filtering processing is:

[0057]

[0058] Among them, g(a, b) represents the Gaussian filtering result of the pixel at position (a, b) in the visual field grayscale image, e represents the exponential base constant, and σ represents the standard deviation of the Gaussian distribution;

[0059] (3.2) Perform histogram equalization processing and contrast enhancement processing on the preprocessed visual field image to obtain the visual field image after image enhancement;

[0060] (3.3) Use the sobel operator to perform edge detection on the visually enhanced field image to obtain the edge features of the visually enhanced field image;

[0061] In this embodiment, the gradient of the Gaussian smoothed visual field image in the horizontal direction is calculated by the Sobel operator convolution kernel in the horizontal direction, the gradient of the Gaussian smoothed visual field image in the vertical direction is calculated by the Sobel operator convolution kernel in the vertical direction, and then the edge features of the visual field image are obtained by calculating the amplitude of the gradients in the two directions.

[0062] (3.4) According to the edge features of the visual field image, use the Hough transform to perform image track straight line detection to obtain the straight line track features of the visual field image;

[0063] In this embodiment, the method of Hough transform straight line detection is as follows: convert the visual field image to the polar coordinate parameter space and discretize it into a grid; according to the edge features of the visual field image, calculate all possible straight lines corresponding to the pixel points with edge features in the visual field image in the parameter space respectively, and perform cumulative counting at the parameter positions of the straight lines; find the pixel points whose cumulative counting results exceed the set threshold, and correspondingly obtain the straight line parameters in the visual field image, so as to obtain the track features of the visual field image.

[0064] (3.5) According to the edge features of the visual field image and the straight line track features of the visual field image, obtain the curve track features of the visual field image through Hough straight line detection and curve fitting;

[0065] (3.5.1) According to the edge features of the visual field image, extend along the straight line track features of the visual field image, and set a trapezoidal region of interest from the visually enhanced field image;

[0066] (3.5.2) Perform perspective transformation on the visually enhanced field image and convert it into a bird's-eye view, and use a cubic function to perform curve fitting within the trapezoidal region of interest according to the straight line track features of the visual field image to obtain the curve track fitting result;

[0067] (3.5.3) If the cubic term coefficient of the curve track fitting result is greater than Then the curve track features of the visual field image are obtained, where is a non-zero constant;

[0068] (3.6) Set the detection limit for track foreign object intrusion;

[0069]

[0070] Among them, u ar (v i' ) represents the right position point of the detection limit for track foreign object intrusion, u ri' represents the characteristic point of the right track of the subway line, u li' represents the characteristic point of the left track of the subway line, u al (v i' ) represents the left position point of the detection limit for track foreign object detection, v i' represents the horizontal coordinate of the i'-th track line;

[0071] (3.7) Replace the standard convolution in the backbone network and feature fusion network of the traditional yolov11 with depthwise separable convolution, then adopt the top-down path aggregation mechanism and lateral connection to strengthen the image feature extraction and analysis between layers of the feature fusion network, and use the C2f-Faster-EMA feature fusion strategy to fuse the image feature information of different scales to be transmitted to the neck network to obtain the improved yolov11 neural network;

[0072] In this embodiment, after replacing the standard convolution in the backbone network and feature fusion network of yolov11 with depthwise separable convolution, since depthwise separable convolution includes depth convolution and point convolution, feature extraction is first performed on each input channel and then channel fusion is performed, which can greatly reduce the number of parameters and computational complexity without affecting the accuracy; in addition, adopting the automatic downward path aggregation mechanism and lateral connection can reduce the additional computational overhead, and using the C2f-Faster-EMA feature fusion strategy can better fuse information of different scales while controlling the computational overhead.

[0073] (3.8) Based on the track line foreign object intrusion dataset, and based on the straight track features of the field of view image, the curved track features of the field of view image, and the detection limit for track foreign object intrusion, perform foreign object detection and recognition training on the improved yolov11 neural network to obtain the foreign object detection neural network for the field of view image;

[0074] In this embodiment, the detection limit for track foreign object intrusion can efficiently determine the region of interest of the detection limit for track foreign object intrusion during foreign object detection and recognition training for monitoring and recognition training. When training the improved yolov11 neural network, half-precision floating-point numbers can be used for gradient update to reduce video memory consumption and improve GPU utilization, thereby accelerating the training speed without affecting convergence.

[0075] (3.9) Use the foreign object detection neural network for the field of view image to detect the real-time field of view image, and obtain the detection result of the field of view image features.

[0076] (4) According to the Kalman filtering algorithm, use the on-vehicle computing device of the subway to identify and monitor the real-time radar data, and obtain the current position of the radar-detected object;

[0077] (4.1) Set a continuous monitoring time window for the real-time radar data in the subway's traveling direction, and count the number of times an object is successfully detected within the continuous monitoring time window as the object detection count;

[0078] (4.2) If the object detection count is less than the first detection count threshold, then determine that the track line state is the foreign object intrusion disappearance state, and return to step (4.1); otherwise, proceed to step (4.3);

[0079] (4.3) If the object detection count is greater than the second detection count threshold, then determine that the track line state is the foreign object intrusion continuous state, and proceed to step (4.4), otherwise directly proceed to step (4.1);

[0080] (4.4) Construct a foreign object intrusion state space model based on the real-time radar data;

[0081]

[0082] v kτ =v k(τ-1) +a k(τ-1) ×Δt

[0083] where p kτ represents the position of the k-th object at time τ, v kτ represents the velocity of the k-th object at time τ, a k(τ-1) represents the acceleration of the k-th object at time τ - 1, and Δt represents the length of the adjacent time interval;

[0084] (4.5) Based on the foreign object intrusion state space model, construct a foreign object intrusion state equation;

[0085]

[0086] where ω τ-1 represents the process noise at time τ - 1, z τ represents the position measurement vector at time τ, H τ-1 represents the observation matrix at time τ - 1, and S τ-1 represents the measurement noise at time τ - 1;

[0087] (4.6) Update the time and state of the foreign object intrusion state equation according to the Kalman filter algorithm to obtain the current position of the object detected by the radar.

[0088] (5) Based on the data synchronization frequency, use the on-vehicle computing device of the subway to fuse the current position of the object detected by the radar and the detection result of the vision image features to obtain the on-vehicle intrusion detection result;

[0089] (5.1) Obtain the data synchronization frequency according to the detection frequency of the radar detection device and the detection frequency of the image detection device;

[0090] (5.2) Obtain the conversion model between the radar coordinate system and the world coordinate system, the conversion model between the pixel coordinate system and the camera coordinate system, and the conversion model between the camera coordinate system and the world coordinate system;

[0091] (5.3) Based on the conversion model between the radar coordinate system and the world coordinate system, convert the current position of the object detected by the radar into the current radar detection position in the world coordinate system;

[0092] (5.4) Based on the conversion model between the pixel coordinate system and the camera coordinate system and the conversion model between the camera coordinate system and the world coordinate system, convert the detection result of the vision image features into the image detection result in the world coordinate system;

[0093] (5.5) According to the data synchronization frequency, perform target position fusion on the current radar detection position and the image detection result at the same moment based on the following target association model to obtain the on-vehicle intrusion detection result;

[0094] Among them, the target association model is:

[0095]

[0096] Among them, Δd represents the distance between the current radar detection position and the center point position of the object in the image detection result, O R represents the current radar detection position, O C represents the center point position of the object in the image detection result, l1 represents the first association degree threshold, IOU represents the intersection over union of the radar and the image detection object regions, IOU represents the contour region of the radar detection object, S C represents the contour region of the image detection object, and l2 represents the second association degree threshold.

[0097] In this embodiment, when the distance between the current radar detection position and the center point position of the object in the image detection result is less than the first association degree threshold, and the intersection over union of the radar and the image detection object regions is greater than the second association degree threshold, the object obtained by the target position fusion can be effectively used as the on-vehicle intrusion detection result.

[0098] (6). The on-vehicle computing device of the subway combines the on-vehicle intrusion detection results and the trackside monitoring foreign object intrusion results, and performs mapping and warning on the subway track line simulation map.

[0099] (6.1) Upload the on-vehicle intrusion detection results to the subway line security center through the on-vehicle communication device of the subway;

[0100] (6.2) Map the trackside monitored foreign object intrusion results and the on-vehicle intrusion detection results to the subway track line simulation map respectively to generate a track foreign object intrusion warning schematic map;

[0101] (6.3) Send the track foreign object intrusion warning schematic map to each on-vehicle communication device of the subway through the subway line security center for warning.

[0102] As Figure 2 shown, the present invention provides a system for identifying subway line foreign object intrusion based on the above-mentioned subway line foreign object intrusion identification method, that is, a subway line foreign object intrusion monitoring system; the subway line foreign object intrusion detection system includes a trackside foreign object intrusion detection subsystem, a subway line security center, and an on-vehicle foreign object intrusion monitoring subsystem that are sequentially arranged along the platforms, curves, and connecting channels of the subway track line. The trackside foreign object intrusion detection subsystem can continuously and comprehensively monitor and identify foreign object intrusion in areas where foreign object intrusion is frequent and prone to occur, and locate and upload the trackside monitored foreign object intrusion results to the subway line security center; the on-vehicle foreign object intrusion monitoring subsystem can monitor and identify foreign object intrusion by combining images and radar, and upload the on-vehicle intrusion detection results in the subway forward direction to the subway line security center; the subway line security center maps the trackside monitored foreign object intrusion results and the on-vehicle intrusion detection results to the subway track line simulation map respectively to generate a track foreign object intrusion warning schematic map, and sends the track foreign object intrusion warning schematic map to each on-vehicle communication device of the subway for warning.

[0103] The present invention comprehensively applies technologies such as trackside monitoring, on-vehicle monitoring, neural network recognition, Kalman filter algorithm, and visualization management, comprehensively improving the detection ability, real-time performance and accuracy, intelligent fusion and target positioning ability, and visualization management and rapid response ability of subway foreign object intrusion, providing a strong guarantee for the operation safety of the subway.

[0104] Although the above-described illustrative specific embodiments of the present invention have been described to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. A method for identifying foreign objects intruding on subway lines, characterized in that: The steps include: (1) Trackside monitoring, positioning, computing and communication devices are installed along the subway track to identify and upload the results of foreign object intrusion by trackside monitoring and to construct a simulated map of the subway track line; (2) Installing an image detection device and a radar detection device on the front of the subway car to capture real-time field of view images in the direction of the subway and to detect and obtain real-time radar data in the direction of the subway; (3) Using the visual field image foreign body detection neural network in the subway vehicle computing device to identify and detect the real-time visual field image, and obtain the visual field image feature detection result; (4) According to the Kalman filter algorithm, the subway vehicle computing device is used to identify and monitor the real-time radar data to obtain the current radar detection object position; (5) According to the data synchronization frequency, the subway vehicle computing device is used to fuse the current radar detection object position and the field of view image feature detection results to obtain the vehicle-mounted intrusion detection results; (6) The subway vehicle computing device combines the on-board intrusion detection results and the trackside monitoring foreign object intrusion results to map the alarm on the subway track line simulation map.

2. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (1) comprises the following steps: (1.1) According to a preset first spacing distance, trackside monitoring devices and trackside computing devices connected to the trackside monitoring devices are sequentially arranged along the platforms, curves and communication passages of the subway track line, wherein the trackside monitoring devices are used to capture trackside monitoring images; and the trackside computing devices are used to identify foreign object intrusion results in the trackside monitoring images; (1.2) According to the preset second spacing distance, trackside positioning devices and trackside communication devices are sequentially arranged along the subway track line, wherein the trackside positioning devices and trackside computing devices at the platform, curve and communication channel are connected to the trackside communication device to transmit the trackside monitoring foreign object intrusion results combined with the line positioning information to the subway line security center; (1.3) Construct a line simulation map according to the subway track line, and map the actual setting positions of the trackside monitoring device, the trackside computing device, the trackside positioning device and the trackside communication device into the line simulation map to obtain a subway track line simulation map.

3. The method for identifying foreign objects intruding into the subway line according to claim 2 is characterized in that: The method for identifying the foreign object intrusion result in the trackside monitoring image in step (1.1) is: (1.1.1) Acquire a number of trackside monitoring images to form a trackside monitoring image dataset; (1.1.2) Keep the backbone network and neck network of the yolov5 network unchanged, replace the three prediction head subnetworks in the prediction head network with the classification prediction subnetwork, the size prediction subnetwork and the offset prediction subnetwork, obtain the trackside image intrusion recognition network, and then set the trackside image intrusion recognition network in the trackside computing device; (1.1.3) According to the multi-scale target loss function, the trackside monitoring image is input into the trackside image intrusion recognition network for intrusion target recognition prediction training to obtain a trained trackside image intrusion recognition network; (1.1.4) Using the trained trackside image intrusion recognition network, the newly captured trackside monitoring image is used to identify intrusion targets, and the original scale intrusion target recognition result is obtained, wherein the original scale intrusion target recognition result includes the position, size and category of each intrusion target; (1.1.5) Determine whether there is an intrusion target whose size is larger than the preset size threshold in the original scale intrusion target recognition result. If so, take the newly captured trackside monitoring image as the large target trackside monitoring image and proceed to step (1.1.6). Otherwise, jump to step (1.1.7); (1.1.6) Set the image scaling factor η s Tune hyperparameters for image resolution; based on image scaling factor The newly captured trackside monitoring image is reduced to obtain a scaled trackside monitoring image; the trained trackside image intrusion recognition network is used to identify intrusion targets in the scaled trackside monitoring image to obtain a scaled intrusion target recognition result; finally, the scaled intrusion target recognition result is combined with the original scaled intrusion target recognition result to obtain the final trackside monitoring foreign object intrusion result; (1.1.7) Set the image scaling factor According to the image scaling factor The newly taken trackside monitoring images are respectively reduced and enlarged to obtain reduced-scale trackside monitoring images and enlarged-scale trackside monitoring images; the trained trackside image intrusion recognition network is used to identify intrusion targets on the reduced-scale and enlarged trackside monitoring images to obtain reduced-scale and enlarged intrusion target recognition results and enlarged-scale intrusion target recognition results; finally, the reduced-scale intrusion target recognition results, the enlarged-scale intrusion target recognition results and the original-scale intrusion target recognition results are unioned to obtain the final trackside monitoring foreign object intrusion result.

4. The method for identifying foreign objects intruding into the subway line according to claim 3 is characterized in that: The multi-scale target loss function is: L=L1+μ1L2+μ2L3 Wherein, L represents the multi-scale target loss function, L1 represents the classification loss function, μ1 represents the first loss function hyperparameter, L2 represents the offset loss function, μ2 represents the second loss function hyperparameter, L3 represents the size loss function, N represents the total number of intrusion targets, C represents the number of channels, H represents the height of the trackside monitoring image, W represents the width of the trackside monitoring image, c represents the category of the pixel in the image, y represents the pixel ordinate of the image, and x represents the pixel abscissa of the image. Indicates the probability that a pixel belongs to the center point of a certain category, υ cyx It represents the true value of the probability that the pixel belongs to the center point of a certain category, α represents the first category exponential factor, β represents the second category exponential factor, represents the offset between the center point of the i-th intrusion target in the trackside monitoring image and the center point of the real target frame in the trackside monitoring image, θ i represents the true value of the offset between the center point of the i-th intrusion target in the trackside monitoring image and the center point of the real target frame in the trackside monitoring image, ||·||1 represents the first norm, represents the size of the i-th intrusion target corresponding to the predicted center point, ξ i Represents the true size of the i-th intrusion target.

5. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (2) comprises the following steps: (2.1) An image detection device is placed at the center of the head of the subway car, and the image detection device is used to capture a real-time field of view image of the subway car's traveling direction; (2.2) Place the radar detection device above the image detection device so that the radar detection device is located on the central axis of the track line; then construct a rectangular coordinate system with the radar detection device as the origin, the subway travel direction as the longitudinal coordinate axis, and the direction perpendicular to the track line as the transverse coordinate axis; (2.3) According to the difference frequency of the Doppler shift, the radar detection device is used to detect the distance, angle and relative speed of the object in the direction of the subway car until the object is detected; (2.4) The detected objects are numbered, and the object numbers are combined with the distance, angle and relative speed corresponding to the objects to obtain real-time radar data of the subway's travel direction.

6. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (3) comprises the following steps: (3.1) Obtain a number of visual field images with or without foreign matter, and perform weighted average grayscale processing and Gaussian filtering processing on the visual field images in sequence to obtain pre-processed visual field images; (3.2) performing histogram equalization and contrast enhancement processing on the preprocessed visual field image to obtain an image-enhanced visual field image; (3.3) Use the Sobel operator to perform edge detection on the visual field image after image enhancement to obtain the edge features of the visual field image; (3.4) According to the edge features of the visual field image, the image track straight line detection is performed using Hough transform to obtain the visual field image straight line track features; (3.5) According to the edge features of the visual field image and the straight track features of the visual field image, the curve track features of the visual field image are obtained by Hough line detection and curve fitting; (3.6) Set the limit for detecting foreign matter intrusion on the track; Among them, u ar (v i' ) represents the right side position point of the track foreign body intrusion detection limit, u ri' Indicates the characteristic point of the right track of the subway line, u li' Indicates the characteristic point of the left track of the subway line, u al (v i' ) represents the left side position point of the track foreign body detection limit, v i' represents the lateral coordinate of the i'th track line; (3.7) Use depthwise separable convolution to replace the standard convolution in the backbone network and feature fusion network of the traditional yolov11, and then use the top-down path aggregation mechanism and lateral connection to strengthen the image feature extraction and analysis between each layer of the feature fusion network, and use the C2f-Faster-EMA feature fusion strategy to fuse image feature information of different scales to transmit to the neck network, and obtain the improved yolov11 neural network; (3.8) According to the track line foreign object intrusion limit data set, based on the field of view image straight track features, field of view image curved track features and track foreign object intrusion limit detection limits, the improved yolov11 neural network is trained for foreign object detection and recognition to obtain the field of view image foreign object detection neural network; (3.9) Use the visual field image foreign body detection neural network to detect the real-time visual field image and obtain the visual field image feature detection result.

7. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (4) comprises the following steps: (4.1) Setting a continuous monitoring time window for the real-time radar data in the direction of subway travel, and counting the number of successful object detections within the continuous monitoring time window as the number of object detections; (4.2) If the number of object detections is less than the first detection number threshold, the track line state is determined to be a foreign object invasion disappearance state, and the process returns to step (4.1); otherwise, the process proceeds to step (4.3); (4.3) If the number of object detections is greater than the second detection number threshold, the track line state is determined to be a foreign object intrusion continuous state, and the process goes to step (4.4); otherwise, the process goes directly to step (4.1); (4.4) Construct a foreign body intrusion state space model based on real-time radar data; in kτ =in k(τ-1) +a k(τ-1) ×Δt Among them, p kτ represents the position of the kth object at time τ, v kτ represents the velocity of the kth object at time τ, a k(τ-1) represents the acceleration of the kth object at time τ-1, and Δt represents the length of the adjacent time interval; (4.5) Based on the foreign body invasion state space model, construct the foreign body invasion state equation; Among them, ω τ-1 represents the process noise at time τ-1, z τ represents the position measurement vector at time τ, H τ-1 represents the observation matrix at time τ-1, S τ-1 represents the measurement noise at time τ-1; (4.6) According to the Kalman filter algorithm, the foreign body intrusion state equation is updated in time and state to obtain the current radar detection object position.

8. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (5) comprises the following steps: (5.1) Obtaining a data synchronization frequency according to the detection frequency of the radar detection device and the detection frequency of the image detection device; (5.2) Obtain the conversion model between the radar coordinate system and the world coordinate system, the conversion model between the pixel coordinate system and the camera coordinate system, and the conversion model between the camera coordinate system and the world coordinate system; (5.3) Based on the conversion model between the radar coordinate system and the world coordinate system, the current radar detection object position is converted into the current radar detection position in the world coordinate system; (5.4) Based on the conversion model between the pixel coordinate system and the camera coordinate system and the conversion model between the camera coordinate system and the world coordinate system, the field of view image feature detection results are converted into image detection results in the world coordinate system; (5.5) According to the data synchronization frequency, the current radar detection position and image detection results at the same time are fused based on the following target association model to obtain the vehicle-mounted intrusion detection result; Among them, the target association model is: Wherein, Δd represents the distance between the current radar detection position and the center point of the object in the image detection result, O R Indicates the current radar detection position, O C represents the center point position of the object in the image detection result, l1 represents the first correlation threshold, IOU represents the intersection-over-union ratio of the radar and image detection object areas, IOU represents the contour area of ​​the radar detection object, S C represents the contour area of ​​the object detected in the image, and l2 represents the second association threshold.

9. The method for identifying foreign objects intruding into the subway line according to claim 1 is characterized in that: The step (6) comprises the following steps: (6.1) Upload the vehicle-borne intrusion detection results to the subway line security center through the subway vehicle-borne communication device; (6.2) Map the foreign object intrusion detection results of the trackside monitoring and the vehicle-mounted intrusion detection results to the subway track line simulation map to generate a track foreign object intrusion warning map; (6.3) Send a warning map of foreign objects intrusion on the track to each subway vehicle's communication device through the subway line security center to issue an alarm.

Images