Train approaching early warning device and method

Abstract

The invention relates to the field of railway traffic safety protection, provides a train approaching early warning method and device, and aims to improve the reliability and flexibility of an early warning system under complex railway working conditions through the combination of an unmanned aerial vehicle, computer vision, image processing and deep learning, help people along the line to effectively avoid dangers and guarantee all-weather operation. Train image data are collected in real time through low-altitude patrol of the unmanned aerial vehicle, then the data are transmitted back to the ground, the train approaching situation is analyzed and judged through an innovative train detection algorithm, and whether an alarm is given or not is decided. By means of the method, intelligent recognition and detection of train approaching are achieved, the reliability, flexibility and accuracy of an early warning system are remarkably improved, and further guarantee is provided for life safety of people along the line.

Description

Technical Field

[0001] The present invention relates to the field of railway traffic safety protection, and in particular to a train approach warning device and method, through which railway staff are warned to ensure their safety. Background Art

[0002] The existing train approach warning methods mainly include traditional manual warning, vehicle-mounted satellite positioning warning and rail-mounted sensor warning. Manual warning means that the railway safety guards monitor and alarm along the line when the railway workers are working. This warning method relies on manual judgment, has a slow response speed, limited coverage, low reliability, many unstable factors, and is easily affected by factors such as the safety officer himself and the environment. Therefore, the accuracy is not high and cannot meet the needs of ensuring the safety of construction personnel. The vehicle-mounted satellite positioning warning method has poor adaptability to complex environments, large data transmission delay at high speeds, and poor anti-interference of wireless signals. The positioning accuracy is low in complex terrain, and it is easy to misreport or miss reports. In addition, it is difficult for the vehicle-mounted system to achieve large-scale coverage and extension, which is not easy to coordinate. For the rail-mounted sensor warning method, its internal structure is complex, the system starts slowly, has poor flexibility, and is complex to deploy. It cannot adapt to mobile operations and is easily affected by environmental factors.

[0003] There are some other early warning methods, but they all have certain application limitations. For example, China Three Gorges University has designed a method for night vehicle recognition using improved YOLOv5, which only considers the situation at night. There is a lack of test data in daytime and more complex weather conditions (rain, fog, snow, etc.). The accuracy and robustness of the model may be low, resulting in a narrowing of the scope of application of this technology. The School of Telecommunications of Jiangsu Information Vocational and Technical College and Binjiang College of Nanjing University of Information Science and Technology jointly designed a YOLOv3 network model based on AlexNet fusion improvement. The weather conditions marked manually may be subjective, especially in weather conditions with blurred boundaries (such as between light fog and cloudy days). This inconsistency may affect the learning effect of the model. Chengdu University of Technology has designed a train arrival warning system based on ZigBee. The sensor module, coordinator module, power supply module, and ZigBee routing submodule of the detection end need to be fixed on the inside of the rail or along the line. The overall flexibility of the system is poor, the disassembly procedure is complicated, and it cannot meet the needs of mobile operations. Summary of the invention

[0004] The present invention comprehensively utilizes the current image acquisition technology, UAV communication technology, image processing technology, target detection technology and deep learning, and proposes a train approach warning method and device for various complex environments. It also innovatively designs a train detection algorithm under complex working conditions by combining UAV, computer vision, image processing and deep learning, and improves the reliability, flexibility and robustness of the warning system, helping personnel along the line to effectively avoid danger and ensure all-weather operations. Help personnel along the line to effectively avoid danger and ensure all-weather operations. The real-time image data of the train is collected and the video data is sent in real time through low-altitude patrol by UAVs, and then the video data sent back is subjected to multiple image processing algorithms and imported into the deep learning model data and transmitted back to the ground. The innovative train detection algorithm is used to analyze and judge whether the train has finally arrived at the train approaching situation and decide whether to alarm. Through this method, the device realizes intelligent image recognition and detection of train approaching, greatly and significantly improves the reliability, flexibility, portability, scope of application and accuracy of the warning system, and further improves and ensures the safety of personnel along the line. Provides further protection for the life safety of personnel along the line.

[0005] A first aspect of the present invention provides a train approach warning device, comprising:

[0006] Image acquisition module, mounted on low-altitude patrol drones, for image acquisition;

[0007] The image processing module and the target detection module are located in the ground server and are used to process the enhanced image data and analyze the approaching train situation and send alarm instructions respectively;

[0008] Image transmission module, used to connect the image acquisition module and the image processing module to ensure efficient transmission of image data stream

[0009] The alarm is carried by the staff and connected to the ground server through the alarm signal transmission module;

[0010] The display screen is connected to the ground server to facilitate real-time manual monitoring.

[0011] As a further improvement, the image acquisition module includes a variety of camera devices and low-visibility fill light components, which acquire images of distant rails at low altitudes, and transmit the image data stream to the image processing module in real time through the image transmission module.

[0012] As a further improvement, the image transmission module includes a UAV ground station / remote controller, an image transmission terminal, and a 5G / 4G network;

[0013] The UAV ground station / remote controller is used to control the flight of the UAV;

[0014] The UAV ground station / remote controller receives the video stream of the image acquisition module and transmits the video stream to the image transmission terminal via the HDMI line;

[0015] The image transmission terminal is used to encode and modulate the video stream, and then send the image data to the image processing module through the 5G / 4G network.

[0016] As a further improvement, the image processing module is loaded with a complex weather impact reduction algorithm. The complex weather impact reduction algorithm first classifies the image according to the degree of impact of different weather on image quality, and then performs adaptive image enhancement, thereby reducing the image noise caused by the main weather.

[0017] As a further improvement, the image processing module is loaded with a complex weather impact reduction algorithm including the following steps: setting day and night as the main impact, whether there is haze as the first secondary impact, and whether it is raining or snowing as the second secondary impact; the original image is first judged by a day and night discriminator. If it is a night image, a dark light enhancement algorithm is used to enhance the illumination and contrast; if it is a daytime image, it enters a haze discriminator for judgment. If there is fog or haze, a dark channel defogging algorithm is used to increase visibility; if there is no fog or haze, it enters a rain and snow discriminator for judgment. If there are raindrops, snowflakes or similar features, a rain and snow removal algorithm is used to reduce rain and snow and reduce noise; if there is no rain or snow, it directly enters the target detection module for identification.

[0018] As a further improvement, the target detection module is based on the YOLOv8 backbone network, and by adding multiple attention mechanisms, the model parameters are modified and adjusted, and then the improved model is verified with a verification set; if the verification is passed, a high-performance train detection model is obtained; if the verification is not passed, training is continued until the verification is passed; after the required model is obtained, the video stream passing through the image processing module is imported into the model for detection to obtain the position and label of the detected target; finally, it is determined whether the detected target exists in the image. If the image detects the target train, the adjacent channel discrimination module and the anti-false detection module are entered to determine whether an alarm instruction needs to be sent, and the detection result is output to the display screen.

[0019] As a further improvement, the adjacent track discrimination module performs image transformation and Canny operator detection on the initial track image taken by the drone to obtain a track boundary image, and then selects a suitable ROI area; due to the high shooting altitude and the small distance between the two rails, the two track lines can be approximated as one track, and then Hough straight line transformation is performed after merging, and then the detected points are linearly fitted, and the position coordinates of the straight area are stored; if it is a partial curve that has not been detected, a small number of reference points are selected in the curve area for rough fitting, and the position coordinates of the curve area are stored after filtering; when the train approaches, an image detected as true by the target detection module is obtained, and the train position reference point is selected to determine whether the reference point is in the track area. If so, it is regarded as a valid image; if not, the next image is directly detected.

[0020] As a further improvement, the anti-false detection module is used to start the image counter when the image is judged to be a valid image, and perform the above-mentioned processing on the first frame, the second frame, ... to the Tth frame image thereafter, and record their validity; the counter counts the number of valid image frames in the T frame image as t, and when t / T is greater than or equal to a certain specific threshold k, an alarm instruction is sent to the alarm signal transmission module, and the counter is reset; if it is less than k, the counter is reset and stopped.

[0021] As a further improvement, the alarm signal transmission module is connected to the target detection module and the alarm. When the target detection module detects a train, a signal is transmitted through the module and received by the worker, and the alarm finally emits a buzzing alarm.

[0022] A second aspect of the present invention provides an early warning method based on the above-mentioned train approach early warning device, comprising the following steps:

[0023] Step 1: When the train approaches the workers' work area, the image acquisition module on the drone hovering low along the track first captures the train video frame, and then sends the video frame to the ground server in real time through the image transmission module. The image processing module processes and enhances the train image in various low-visibility environments into high-quality, low-noise, and easy-to-detect images;

[0024] Step 2: Enter the target detection module and import the enhanced image into the target detection improved model with multiple attention mechanisms.

[0025] If the model detects a train and it is coming from the same track or adjacent track, after judging that it is not a false detection, it will send out an alarm signal through the alarm signal transmission module. The alarm device on the worker will receive the signal and send out an alarm until the train leaves the drone's field of view, the alarm stops, and the device enters the standby state for the next alarm.

[0026] If no train is detected, no alarm is given.

[0027] The application of the technical solution of the present invention has the following technical effects:

[0028] Through the present invention, in complex working conditions (night, rain, snow, haze, dust, etc.), when the railway maintenance personnel are repairing the rails, when the train approaches the workers' maintenance area, the image acquisition module on the drone patrolling along the track at low altitude captures the train video frame and transmits it to the image processing module of the server in real time. The image processing module of the server (including complex weather impact reduction algorithm, etc.) enhances the image processing in various low-visibility environments into high-quality, low-noise, and easy-to-detect images. Then enter the target detection module to analyze and judge the approaching situation of the train. If the target train is detected and it is a train coming from this or adjacent lanes and there is no false detection, an alarm instruction is issued, and the alarm device on the worker's body alarms, indicating the worker to evacuate. In this way, through the low-altitude patrol and mobile flexibility of the drone and the fast and efficient image processing and target detection technology, it can provide comprehensive and highly reliable early warnings in various complex working conditions (such as night, rain, snow, haze, etc.). By early warning and giving enough evacuation time, the present invention can effectively reduce the incidence of train collision accidents caused by untimely avoidance, avoid danger in time, improve maintenance efficiency, and guard the safety of operators. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0030] Figure 1 A schematic diagram showing the appearance of the early warning device of the present invention is shown;

[0031] Figure 2 The overall structural intent of the early warning device of the present invention is shown;

[0032] Figure 3 A schematic diagram of an image transmission module of the early warning device of the present invention is shown;

[0033] Figure 4 A schematic diagram showing an algorithm for reducing the impact of complex weather in an image processing module of an early warning device of the present invention is shown;

[0034] Figure 5 A schematic diagram showing a target detection module in an image processing module of an early warning device of the present invention is shown;

[0035] Figure 6 A schematic diagram showing a local adjacent channel discrimination module in an image processing module of the early warning device of the present invention is shown;

[0036] Figure 7 A schematic diagram of the anti-false detection module in the image processing module of the early warning device of the present invention is shown. DETAILED DESCRIPTION

[0037] The following will describe the implementation methods of the present application in detail with the help of accompanying drawings and examples, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0038] Example 1

[0039] The first embodiment of the present invention provides a train approach warning device, the overall structure of which is as follows: Figure 1-2 As shown, the device includes: an image acquisition module, an image processing module, a target detection module, an image transmission module, an alarm signal transmission module (transmitter, receiver, etc.), an alarm and a display screen.

[0040] The image acquisition module is mounted on a low-flying drone, the image processing module and the target detection module are located on the ground server and are connected by an image transmission module; the alarm is located on the worker (safety helmet, work clothes, etc.) and is connected to the ground server by an alarm signal transmission module.

[0041] When the train approaches the workers' work area, the image acquisition module on the drone hovering low along the track first captures the video frame of the train, and then sends the video frame to the ground server in real time through the image transmission module. The image processing module of the server enhances the image processing in various low-visibility environments into high-quality, low-noise, and easy-to-detect images. Then enter the target detection module, import the enhanced image into the target detection improved model with multiple attention mechanisms, if the model detects the train and it is a train coming from the adjacent lane, after judging that there is no false detection, the alarm signal transmission module will send an alarm signal, and the alarm on the worker will receive the signal and send an alarm (whistle, vibration, etc.) until the train leaves the drone's field of view, the alarm stops, and the device enters the standby state for the next alarm; if the train is not detected, no alarm will be given.

[0042] The image acquisition module includes multiple camera devices such as RGB cameras, IR cameras and low-visibility fill light components, which can flexibly handle image shooting in various complex environments. The image acquisition module is mounted on the drone, collects images of distant rails at a high altitude, and transmits the image data stream to the image processing module in real time through the image transmission module.

[0043] like Figure 3As shown, the image transmission module is connected to the image acquisition module and the image processing module, and is responsible for receiving the video stream of the image acquisition module and sending it to the image processing module. The image transmission module includes a UAV ground station / remote controller, an image transmission terminal, and a 5G / 4G network. The UAV ground station / remote controller controls the image acquisition by controlling the flight of the UAV. The image acquisition module transmits the video stream to the UAV ground station / remote controller. The UAV ground station transmits the video stream to the image transmission terminal via the HDMI line. After encoding and modulating the video stream, the image transmission terminal sends the video signal to the image processing module via the 5G / 4G network.

[0044] like Figure 4 As shown, in the image processing module, the present invention proposes a complex weather impact reduction algorithm. According to the degree of influence of different weather on image quality, the image is first classified by weather, and then adaptive image enhancement is performed to reduce the impact of the main weather. We set day and night as the main influence, whether there is haze as the first secondary influence, and whether it is raining or snowing as the second secondary influence; the original image is first judged by the day and night discriminator. If it is a night image, the dark light enhancement algorithm (including but not limited to MSRCR, MSRCP, ret i nexnet algorithm, LI ME algorithm, MBLLEN algorithm, etc.) is used to enhance the illumination and contrast; if it is a daytime image, it enters the haze discriminator for judgment. If there is fog or haze, the dark channel defogging algorithm is used to increase visibility; if there is no fog or haze, it enters the rain and snow discriminator for judgment. If there are raindrops, snowflakes or similar features, the rain and snow removal algorithm is used to reduce rain and snow and reduce noise; if there is no rain or snow, it directly enters the target detection module for recognition. After the processed images of each path are recognized by the target detection module, if the result is false, the process is stopped and the above operation is repeated for the next image; if it is true, the adjacent track discrimination module is entered to determine the track where the oncoming vehicle is located; if it is a distant track, it is stopped; if it is the current track or an adjacent track, the frame image is regarded as a valid image and enters the anti-false detection module; after the anti-false detection module passes, an alarm command is sent to alarm.

[0045] This algorithm improves the recognition accuracy of trains in various weather conditions by weakening the main weather factors. It does not require multiple deep learning processes and avoids the problems of complex network structure and low time efficiency caused by training detection models in different weather conditions in existing algorithms. The algorithm has a simple structure, is efficient and fast, and meets the real-time and low-latency requirements of early warning devices.

[0046] The structure of the target detection module is as follows: Figure 5As shown in the figure, this module is based on the YOLOv8 backbone network and achieves higher performance target detection in complex weather by adding multiple attention mechanisms (including CBAM, GAM, etc.). First, the target detection model is trained with the preprocessed complex weather training set, the internal parameters of the model are modified, and then the improved model is verified with the preprocessed verification set. If the verification is passed, a target detection model with high recognition accuracy is obtained; if the verification is not passed, the training is continued until the verification is passed to obtain a target detection model with high recognition accuracy. After obtaining the required model, the video stream that has passed the image processing module (complex weather impact reduction algorithm, etc.) is imported into the model for detection to obtain the location and label of the detected target. Finally, it is determined whether the detected target exists in the image. If the image detects the target train, it enters the adjacent channel discrimination module and the anti-false detection module to determine whether an alarm instruction needs to be sent, and outputs the detection result to the display screen.

[0047] The structure of this adjacent channel discrimination module is as follows Figure 6 As shown in the figure. After the early warning device is activated, the UAV flies along the track to an open area and a point where the track line is close to a straight line, then hovers, and the camera is stationary facing the rail to shoot the initial track image. The initial track image is transformed by image transformation (Gray, Gaussian Blur, etc.) and Canny operator detection to obtain the track boundary image, and then a suitable ROI area is selected (because the shooting angle makes the track in the middle of the image, the area with constant height and width from three-eighths to five-eighths is temporarily selected); because the shooting height is high and the distance between the two rails is small, the two track lines can be approximated as one track, and Hough Lines are performed after merging, and then the detected points are linearly fitted, and then some straight lines are filtered according to the slope and the appropriate safety area is expanded, and then the position coordinate information of the straight area is stored; if it is a partial curve that is not detected, a small number of reference points are selected in the curve area for rough linear fitting, and the position coordinate information of the curve area is stored after filtering. When the train approaches, the target detection module obtains the image detected as true, selects the train position reference point, and determines whether the reference point is within the track area. If so, it is considered a valid image; if not, the next image is directly detected.

[0048] The structure of the anti-false detection module is as follows: Figure 7As shown. When the image is judged as a valid image, the image counter is turned on, and the above processing is performed on the first frame, the second frame, ... to the Tth frame (the interval can be other integers, which is set to 1 here) and their validity is recorded. The counter counts the number of valid image frames in the T-frame image as t. When t / T is greater than or equal to a certain threshold k, an alarm instruction is sent to the alarm signal transmission module and the counter is reset; if it is less than k, the counter is reset and stopped. Among them, the selection of T, t, and k is the result of comprehensive consideration of factors such as train approach time, FPS, and actual test data.

[0049] The alarm signal transmission module is connected to the target detection module and the alarm. When the target detection module detects a train, a signal is transmitted through the module, which is received by the worker and the alarm finally sounds.

[0050] The alarm is controlled by the alarm signal transmission module, and sounds an alarm when the alarm signal transmission module issues an instruction to it.

[0051] The processed detection-enhanced video shot by the drone is displayed on the display screen, which facilitates manual assistance in monitoring and alarm.

[0052] Example 2

[0053] The second embodiment of the present invention provides an early warning method using the first embodiment. When a train approaches a worker's work area, an image acquisition module on a drone hovering at low altitude along the track first captures a video frame of the train, and then sends the video frame to a ground server in real time through an image transmission module. The image processing module of the server can process and enhance images in various low-visibility environments into high-quality, low-noise, and easy-to-detect images; then enter the target detection module, and import the enhanced image into an improved target detection model that adds multiple attention mechanisms; if the model detects a train and it is a vehicle coming from the adjacent lane, an alarm signal is sent out through an alarm signal transmission module after it is determined that there is no false detection, and an alarm device on the worker receives the signal and sends an alarm until the train leaves the drone's field of view, the alarm stops, and the device enters a standby state for the next alarm; if no train is detected, no alarm is given.

[0054] The following is an example.

[0055] Case 1

[0056] The train approach warning device can warn people along the line of approaching trains in various complex working conditions to ensure their safety. After the people along the line arrive at the work site, the device is started, and the drone flies along the track at a suitable height to an open area or a place where the track line is close to a straight line. The drone turns on the camera to shoot and transmits the video stream to the ground server. The server classifies the weather according to the image features of the video frame, and then uses different image processing algorithms to enhance the video frame according to the degree of weather influence. The target detection module identifies the video frame, confirms that it is an approaching vehicle in the adjacent lane and there is no false detection, and sends the identification result and alarm command to the display screen and alarm respectively. After receiving the command, the alarm sounds an alarm, and the people along the line avoid it in time. The video frames in different working conditions have different image features and different processing methods.

[0057] If the construction is carried out in the evening or at night when the lighting is dim, and the brightness and contrast of the captured video frame do not reach the threshold, the image is judged as night, and the image is enhanced in low light before target detection. If the target is detected and confirmed to be a vehicle on the same track or adjacent track and there is no false detection, an alarm command is sent to stop processing the video frame.

[0058] If the construction is carried out in low visibility scenes such as haze, sand and dust during the day, the video frame is detected as daytime and then enters the next level to determine whether it is foggy. If it is foggy, it will be subjected to dark channel defogging and then target detection. If the target is detected and confirmed to be a vehicle on the same track or adjacent track and there is no false detection, an alarm command will be sent to stop processing the video frame.

[0059] If construction is carried out in a scene with high noise such as rain or snow, the video frame shot will enter the final level of rain and snow judgment after being detected as daytime and fog-free. If there is rain or snow, it will be processed for rain and snow removal, and then target detection will be performed. If the target is detected and confirmed to be a vehicle on the same track or adjacent track and there is no false detection, an alarm command will be sent to stop processing the video frame.

[0060] If the construction is carried out in normal weather such as daytime or cloudy days, and the video frame is detected as daytime, fog-free, rain-free or snow-free, target detection is performed directly. If the target is detected and confirmed to be a vehicle on the same track or adjacent track and there is no false detection, an alarm command is sent to stop processing the video frame.

[0061] Case 2

[0062] When railway maintenance personnel are working in mountainous railway sections with complex terrain, numerous tunnels and bridges, the sight is easily obstructed when trains pass through these areas. Traditional early warning systems have limited effectiveness in these complex environments and cannot effectively warn of upcoming trains, making it difficult for workers to evacuate in time in emergencies, posing a major safety hazard. Our system can solve this problem.

[0063] Flexible early warning system: During each inspection, the device will use a drone to cruise in the air at a distance from the inspection area. The drone's camera and sensors monitor in real time whether there is a train approaching on the railway. The drone is flexible in flight, has a wide field of view, and is easy to carry, thus eliminating the impact of complex terrain.

[0064] Real-time warning: When the system detects a train approaching, the device will immediately send an alert to workers in the construction area. The alert is sent through the ground server and synchronously transmitted to the smart devices worn by the workers, ensuring that they can evacuate to a safe area in time.

[0065] Environmental adaptability: The images taken by the drone are enhanced for complex weather environments such as night, fog, rain, snow, and dust, thereby comprehensively enhancing the detection accuracy of trains. Even in adverse weather conditions, the drone can maintain stable flight and monitoring and provide accurate warning information.

[0066] If the brightness of the video frame does not reach the threshold, the image is judged as dark, and the image is enhanced for dark light, and then the target is detected. If the target is detected, an alarm command is sent and the processing of the video frame is stopped; if the target is not detected, the video frame is judged whether it is foggy.

[0067] If the video frame is detected as foggy, dark channel defogging is performed on it, and then target detection is performed. If a target is detected, an alarm command is sent and the processing of the video frame is stopped; if no target is detected, the processing of the video frame is stopped.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A train approach warning device, characterized in that: include: Image acquisition module, mounted on low-altitude patrol drones, for image acquisition; The image processing module and the target detection module are located in the ground server and are used to process the enhanced image data and analyze the approaching train situation and send alarm instructions respectively; Image transmission module, used to connect the image acquisition module and the image processing module to ensure efficient transmission of image data stream The alarm is carried by the staff and connected to the ground server through the alarm signal transmission module; The display screen is connected to the ground server to facilitate real-time manual monitoring.

2. The train approach warning device according to claim 1, characterized in that: The image acquisition module includes a variety of camera equipment and low-visibility fill light components, which collect images of distant rails at low altitudes, and transmit image data streams to the image processing module in real time through the image transmission module.

3. The train approach warning device according to claim 2, characterized in that: The image transmission module includes a UAV ground station / remote controller, an image transmission terminal, and a 5G / 4G network; The UAV ground station / remote controller is used to control the flight of the UAV; The drone ground station / remote controller receives the video stream from the image acquisition module and transmits the video stream to the image transmission terminal via an HDMI cable; The image transmission terminal is used to encode and modulate the video stream, and then send the image data to the image processing module through the 5G / 4G network.

4. The train approach warning device according to claim 3, characterized in that: The image processing module is loaded with a complex weather impact reduction algorithm. The complex weather impact reduction algorithm first classifies the image according to the degree of impact of different weather on image quality, and then performs adaptive image enhancement, thereby reducing the image noise of the main weather.

5. The train approach warning device according to claim 4, characterized in that: The image processing module is loaded with a complex weather impact reduction algorithm, which includes the following steps: setting day and night as the main impact, whether there is haze as the first secondary impact, and whether it is raining or snowing as the second secondary impact; the original image is first judged by a day and night discriminator, and if it is a night image, a dark light enhancement algorithm is used to enhance the illumination and contrast; if it is a daytime image, it enters the haze discriminator for judgment, and if there is fog or haze, a dark channel defogging algorithm is used to increase visibility; if there is no fog or haze, it enters the rain and snow discriminator for judgment, and if there are raindrops, snowflakes or similar features, a rain and snow removal algorithm is used to reduce rain and snow and reduce noise; if there is no rain or snow, it directly enters the target detection module for identification.

6. The train approach warning device according to claim 3, characterized in that: The target detection module is based on the YOLOv8 backbone network. By adding multiple attention mechanisms, the model parameters are modified and adjusted, and the improved model is verified with a verification set. If the verification is passed, a high-performance train detection model is obtained. If the verification is not passed, training is continued until the verification is passed. After the required model is obtained, the video stream passing through the image processing module is imported into the model for detection to obtain the position and label of the detected target. Finally, it is determined whether the detected target exists in the image. If the image detects the target train, the adjacent channel discrimination module and the anti-false detection module are entered to determine whether an alarm instruction needs to be sent, and the detection result is output to the display screen.

7. The train approach warning device according to claim 6, characterized in that: The adjacent track discrimination module performs image transformation and Canny operator detection on the initial track image taken by the drone to obtain the track boundary image, and then selects the appropriate ROI area; due to the high shooting height and the small distance between the two rails, the two track lines can be approximated as one track, and the Hough straight line transformation is performed after merging, and then the detected points are linearly fitted, and the position coordinates of the straight area are stored; if it is a partial curve that has not been detected, a small number of reference points are selected in the curve area for rough fitting, and the position coordinates of the curve area are stored after filtering; when the train approaches, an image detected as true by the target detection module is obtained, and the train position reference point is selected to determine whether the reference point is in the track area. If so, it is regarded as a valid image; if not, the next image is directly detected.

8. The train approach warning device according to claim 7, characterized in that: The anti-false detection module is used to start the image counter when the image is judged to be a valid image, perform the above processing on the first frame, the second frame, ... to the T-th frame image thereafter, and record their validity; The counter counts the number of valid image frames in T frame images as t. When t / T is greater than or equal to a certain threshold k, an alarm instruction is sent to the alarm signal transmission module and the counter is reset; if it is less than k, the counter is reset and stopped.

9. The train approach warning device according to claim 1, characterized in that: The alarm signal transmission module is connected to the target detection module and the alarm. When the target detection module detects a train, a signal is transmitted through the module and received by the worker, and the alarm finally emits a buzzing alarm.

10. The early warning method of the train approach early warning device according to any one of claims 1 to 9, comprising the following steps: Step 1: When the train approaches the workers' work area, the image acquisition module on the drone hovering low along the track first captures the train video frame, and then sends the video frame to the ground server in real time through the image transmission module. The image processing module processes and enhances the train image in various low-visibility environments into high-quality, low-noise, and easy-to-detect images; Step 2: Enter the target detection module and import the enhanced image into the target detection improved model with multiple attention mechanisms. If the model detects a train and it is coming from the same track or adjacent track, after judging that it is not a false detection, it will send out an alarm signal through the alarm signal transmission module. The alarm device on the worker will receive the signal and send out an alarm until the train leaves the drone's field of view, the alarm stops, and the device enters the standby state for the next alarm. If no train is detected, no alarm is given.

Images