Obstacle detection method and system for rail transit vehicle

Abstract

The invention discloses an obstacle detection method and system for a rail transit vehicle. The method is applied to the technical field of rail transit safety. The method comprises the steps that various types of driving environment data in the vehicle driving process are collected in real time; based on a preset driving environment prediction model, generating a virtual vision field representing a predicted driving environment within a beyond-visual-range range by using the actual driving environment data within the visual-range range; the actual driving environment data in the sight distance range and the environment data belonging to the obstacle in the virtual vision field are marked, and then based on a preset obstacle detection model, the driving environment data carrying the mark are utilized to generate a corresponding real-time obstacle detection result. According to the invention, the dynamic obstacle detection precision in the vehicle driving process is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of rail transit safety technology, and in particular relates to a method and system for detecting obstacles for rail transit vehicles. BACKGROUND

[0002] Currently, obstacle detection and early warning in rail transit systems mainly rely on the cooperation of on-board and wayside sensors, and the detection and monitoring of obstacles are realized through vehicle-ground cooperation, which has certain limitations. For example: A Chinese invention patent with publication number CN116605268B discloses a vehicle-ground cooperative perception mobile authorization calculation method, relating to the field of rail train signal control technology. The method includes the steps of: obtaining the ID of the Starlink TSL that needs to communicate; finding the ID of the Starlink TSL at the end point of the ITE sensing range of the Eagle Eye system, and splicing the sensing ranges of the Eagle Eye system ITE and the Starlink TSL; obtaining the position of the obstacle and performing obstacle screening; calculating the Starlink TSL sensing recognition point and the sensing recognition state; and performing Starlink TSL mobile authorization calculation according to the sensing recognition point and the sensing recognition state. The Eagle Eye system ITE of the present application communicates with the Starlink TSL, splices the sensing ranges continuously, and perceives the sensing recognition of obstacle information, thereby preventing faults that have occurred in front of the vehicle earlier. This method can obtain the position information of the obstacle in front of the vehicle and perform screening through the splicing of the sensing ranges of the Eagle Eye system and the Starlink, thereby achieving earlier fault prevention. However, this method relies on the precise communication and data synchronization of the Starlink system, and has weak recognition and prediction capabilities for rapidly changing obstacles in dynamic environments, especially the detection accuracy and reaction time for obstacles beyond the sensing range still need to be improved.

[0003] A Chinese invention patent with publication number CN116654053A discloses a rail transit vehicle-ground cooperative beyond-visual-range obstacle detection system and method, which includes a vehicle-mounted obstacle detection host and obstacle detection sensors, trackside obstacle detection sensors, and a CBTC signal system. During train operation in full automatic driving mode, the vehicle-mounted obstacle sensors are responsible for in-visual-range obstacle detection and upload the detection results to the vehicle-mounted obstacle detection host, and the trackside sensors are responsible for out-visual-range obstacle detection and upload the results to the vehicle-mounted obstacle detection host through TWC wireless communication. The present invention detects obstacles within and outside the visual range in front of the train through vehicle-mounted obstacle sensors, trackside sensors, and a vehicle-mounted obstacle detection host, uploads the detection information to the vehicle-mounted ATC and trackside ATC, CBI, ATS, and integrated monitoring subsystems after intelligent operation and judgment, realizes reliable obstacle detection function during train operation, greatly improves train operation safety protection performance and operation efficiency. The system realizes obstacle detection in full automatic driving mode through the vehicle-mounted obstacle detection host and obstacle detection sensors, combined with trackside obstacle detection sensors and signal systems, can detect obstacles within and outside the visual range of the train, and further improves train operation safety by transmitting data to the vehicle-mounted host through wireless communication. However, the system relies on data integration and transmission of multiple sensors, which may have data transmission delay, sensor error, and limitations on beyond-visual-range obstacle detection, especially weak tracking and trajectory prediction ability for dynamic obstacles. SUMMARY

[0004] One of the technical problems to be solved by the present application is to provide an obstacle detection method for rail transit vehicles, which comprises: real-time collection of various types of driving environment data during vehicle driving; based on a preset driving environment prediction model, using actual driving environment data within the visual range to generate a virtual visual field representing the predicted driving environment within the beyond-visual-range; labeling the actual driving environment data within the visual range and the environment data belonging to obstacles in the virtual visual field, and then based on a preset obstacle detection model, using the labeled driving environment data to generate corresponding real-time obstacle detection results.

[0005] In combination with the above implementation modes, before generating a virtual visual field representing a predicted driving environment within a beyond-visual-range based on a preset driving environment prediction model using actual driving environment data within a visual range, the obstacle detection method further comprises: The preset first deep learning model is trained based on historical actual driving environment data within the line-of-sight range and the line-of-sight range during vehicle driving in a historical period, so as to obtain a preset driving environment prediction model. The preset driving environment prediction model is composed of a PointNet++ and a convolutional neural network framework.

[0006] In combination with the above implementation manners, before the obstacle detection method generates the real-time obstacle detection result based on the preset obstacle detection model and the driving environment data carrying the label, the obstacle detection method further comprises: The environment data belonging to the obstacle in the historical actual driving environment data within the line-of-sight range and the line-of-sight range is labeled, and then the preset second deep learning model is trained based on the historical driving environment data carrying the label as the input information of the preset second deep learning model and the environment data belonging to the obstacle as the output information of the preset second deep learning model, so as to obtain the preset obstacle detection model. The preset second deep learning model is composed of a Transformer variant framework based on a self-attention mechanism.

[0007] In combination with the above implementation manners, the types of the driving environment data include but are not limited to point cloud data and video stream data, and the obstacle detection method further comprises: The YOLOv5, 3DSSD and DeepSORT algorithms are added to the preset second deep learning model to obtain the spatial position change characteristics of each obstacle in a continuous period, so as to realize the intention prediction and trajectory prediction of each obstacle.

[0008] In combination with the above implementation manners, the obstacle detection method further comprises: According to the generated real-time obstacle detection result, the collision time and the collision distance between the obstacle and the vehicle are predicted. According to the collision time prediction result and the collision distance prediction result respectively, the first collision risk degree and the second collision risk degree are generated, and then the comprehensive collision risk degree is obtained by analyzing and comparing the collision risk degrees, and a warning is given.

[0009] In combination with the above implementation manners, the obstacle detection method further comprises: The system uses a multi-head attention mechanism to monitor key data that affects the reliability of early warnings in real time, and adopts a parallel processing approach for different types of key data to assess global risks. Then, it uses the results of the global risk assessment to verify the current level of risk and issue early warnings based on the verified level of risk.

[0010] Combining the above implementation methods, the global risk is assessed using the following expression: ; in, This indicates the overall risk assessment result. n Indicates the number of attention heads. Indicates the first i The weighting coefficient of each attention head. Indicates the first i The output of each attention head, , , and These represent the first [item] in different types of key data. i Data in multiple dimensions.

[0011] In conjunction with the above implementation methods, before generating a virtual field of view representing the predicted driving environment beyond the line-of-sight range, the obstacle detection method further includes: All driving environment data is aligned to a unified spatiotemporal coordinate system to obtain the best virtual field of view.

[0012] In conjunction with the above implementation methods, obstacle detection methods also include: Point cloud data is collected using vehicle-mounted LiDAR deployed on the vehicle; Video stream data is collected by ground-based binocular cameras deployed in a ground area along the vehicle's trajectory.

[0013] On the other hand, this application also provides an obstacle detection system for rail transit vehicles, the system comprising the following modules: The data processing module is used to collect various types of driving environment data in real time during vehicle operation; The driving environment prediction module is used to generate a virtual field of view representing the predicted driving environment beyond the line of sight, based on a preset driving environment prediction model and using actual driving environment data within the line of sight range. The obstacle detection module is used to mark the actual driving environment data within the line of sight and the environmental data of obstacles in the virtual field of view. Then, based on the preset obstacle detection model, it uses the marked driving environment data to generate the corresponding real-time obstacle detection results.

[0014] Compared with the prior art, one or more embodiments in the above scheme can have the following advantages or beneficial effects: The application provides a method and system for obstacle detection of a rail transit vehicle. The method collects various types of driving environment data in real time during vehicle driving based on vehicle-ground cooperation. Then, based on a deep learning algorithm, a preset driving environment prediction model and a preset obstacle detection model are trained by using historical actual driving environment data and environment data belonging to obstacles in a historical period, and real-time obstacle detection results are obtained on this basis. The application effectively improves the detection accuracy of dynamic obstacles during vehicle driving, can accurately identify obstacles within the range of beyond visual range during the operation of the rail transit vehicle, and provides technical support for reducing accidents caused by obstacles and reducing downtime and maintenance costs caused by accidents. With the improvement of the intelligent level of rail transit, the market demand for related equipment and technology will continue to grow. The application effectively improves the safety of the rail transit system while improving the transportation efficiency, and will have a wide application prospect in the future (especially in the research and development of intelligent rail transit systems and automatic driving trains), and has high market value.

[0015] Other features and advantages of the application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the application. The objects and other advantages of the application can be achieved and obtained by the structure specifically pointed out in the specification and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are intended to provide a further understanding of the application and constitute a part of the specification, which together with the embodiments of the application, serve to explain the application, and do not constitute a limitation on the application. In the drawings: Figure 1 A flowchart of the method for obstacle detection of a rail transit vehicle according to an embodiment of the application.

[0017] Figure 2 A module block diagram of the system for obstacle detection of a rail transit vehicle according to an embodiment of the application. DETAILED DESCRIPTION

[0018] The embodiments of the application will be described in detail below with reference to the accompanying drawings and embodiments, so that how the application applies technical means to solve technical problems and achieves technical effects can be fully understood and implemented. It should be noted that, unless there is a conflict, each embodiment in the application and each feature in the embodiment can be combined with each other, and the formed technical scheme is within the protection scope of the application.

[0019] Additionally, the steps shown in the flowcharts of the drawings can be performed in a computer system such as a set of computer-executable instructions, and while a logical sequence is shown in the flowcharts, in some cases the steps shown or described can be performed in an order other than that shown.

[0020] Currently, obstacle detection and early warning in rail transit systems mainly rely on the cooperation of on-board and wayside sensors, and the detection and monitoring of obstacles are realized through vehicle-ground cooperation, which has certain limitations. For example: The Chinese invention patent with publication number CN116605268B discloses a vehicle-ground cooperative perception mobile authorization calculation method, relating to the field of rail train signal control technology. It includes the steps of: obtaining the ID of the Starlink TSL that needs to communicate; finding the ID of the Starlink TSL at the end of the ITE perception range of the Eagle Eye system, and splicing the perception ranges of the Eagle Eye system ITE and the Starlink TSL; obtaining the position of the obstacle and performing obstacle screening; calculating the Starlink TSL perception recognition points and the Starlink TSL perception recognition state; and performing Starlink TSL mobile authorization calculation according to the perception recognition points and the Starlink TSL mobile authorization calculation according to the perception recognition state. The Eagle Eye system ITE of the present invention communicates with the Starlink TSL, splices the range of view continuously, and perceives the obstacle information and the like, thereby preventing faults that have occurred in front of the vehicle earlier. This method can obtain the position information of the obstacle in front of the vehicle and perform screening through the splicing of the perception ranges of the Eagle Eye system and the Starlink, thereby achieving earlier fault prevention. However, this method relies on the precise communication and data synchronization of the Starlink system, and has weak recognition and prediction ability for fast-changing obstacles in dynamic environments, especially the detection accuracy and reaction time for obstacles beyond the range of view still need to be improved.

[0021] A Chinese invention patent with publication number CN116654053A discloses a rail transit vehicle-ground cooperative beyond-visual-range obstacle detection system and method, which includes a vehicle-mounted obstacle detection host and obstacle detection sensors, trackside obstacle detection sensors, and a CBTC signal system. During train operation in full automatic driving mode, the vehicle-mounted obstacle sensors are responsible for obstacle detection within the sensor's visual range and upload the detection results to the vehicle-mounted obstacle detection host, and the trackside sensors are responsible for obstacle detection outside the vehicle-mounted sensor's visual range and upload the results to the vehicle-mounted obstacle detection host through TWC wireless communication. The present invention detects obstacles within and outside the visual range in front of the train through vehicle-mounted obstacle sensors, trackside sensors, and the vehicle-mounted obstacle detection host, uploads the detection information to the vehicle-mounted ATC and the trackside ATC, CBI, ATS, and integrated monitoring subsystem after intelligent operation and judgment, realizes reliable obstacle detection function during train operation, and greatly improves train operation safety protection performance and operation efficiency. The system realizes obstacle detection in full automatic driving mode through the vehicle-mounted obstacle detection host and obstacle detection sensors, combined with trackside obstacle detection sensors and signal systems, can detect obstacles within and outside the train's visual range, and further improves train operation safety by transmitting data to the vehicle-mounted host through wireless communication. However, the system relies on data integration and transmission of multiple sensors, which may have data transmission delay, sensor error, and limitations on beyond-visual-range obstacle detection, especially weak tracking and trajectory prediction ability for dynamic obstacles.

[0022] Therefore, to solve the above problems, the present application provides an obstacle detection method and system for rail transit vehicles. The method collects various types of driving environment data in real time during vehicle operation based on vehicle-ground cooperation. Then, based on deep learning algorithm, the historical actual driving environment data and the environment data belonging to obstacles in the historical driving process of the vehicle in the historical period are used to train a preset driving environment prediction model and a preset obstacle detection model, and the corresponding real-time obstacle detection result is obtained. The present application effectively improves the detection accuracy of dynamic obstacles during vehicle operation, can accurately identify obstacles within the beyond-visual-range range during rail transit vehicle operation, and provides technical support for reducing accidents caused by obstacles and reducing downtime and maintenance costs caused by accidents. With the improvement of the intelligent level of rail transit, the market demand for related equipment and technology will continue to grow, and the present application effectively improves the safety of the rail transit system while improving the transportation efficiency, and will have a wide application prospect in the future (especially in the research and development of intelligent rail transit systems and automatic driving trains), and has high market value.

[0023] Example One Figure 1A flowchart of the steps of the obstacle detection method for rail transit vehicles according to an embodiment of the present application. The various steps of the method will be described below with reference to Figure 1 FIG. 1.

[0024] As shown in FIG. 1, in step S110, various types of driving environment data are collected in real time during vehicle driving. Figure 1

[0025] Specifically, the various types of driving environment data are collected in real time during vehicle driving in the embodiment. After the driving environment data are collected, the various types of driving environment data are preprocessed by using filtering, denoising, enhancement, and other preprocessing methods, and the preprocessed driving environment data belonging to the same time are determined, thereby obtaining the basic data for obstacle detection.

[0026] Further, the collected driving environment data are collected in real time by the data transmission unit, and the collected driving environment data are transmitted in real time and efficiently to the monitoring center through 4G, 5G, and other wireless networks, thereby realizing real-time communication between the vehicle and the ground and remote communication, ensuring information synchronization between the vehicle and the ground, and facilitating remote monitoring and management. After the real-time driving environment data are received by the monitoring center in the first time, further data analysis is performed by the cloud platform, thereby providing technical support for remote decision-making and collaborative work for obstacle detection, and enabling the present application to support more intelligent decision-making and remote intervention.

[0027] Further, before the virtual visual field representing the predicted driving environment in the beyond-visual-range is generated, all the driving environment data are aligned to a unified space-time coordinate system to obtain the best virtual visual field.

[0028] In the embodiment of the present application, all the driving environment data are aligned to the unified space-time coordinate system by time stamp synchronization and coordinate transformation alignment processing between different types of driving environment data. The driving environment data aligned to the unified space-time coordinate system can obtain a virtual visual field that is more consistent with the actual driving environment, and the virtual visual field is the best virtual visual field. Therefore, the present application ensures the consistency of driving environment data from different sources in space and time, and enables subsequent analysis and utilization of the driving environment data to be effectively performed.

[0029] Further, in step S120, based on a preset driving environment prediction model, the virtual visual field representing the predicted driving environment in the beyond-visual-range is generated by using the actual driving environment data in the visual range.

[0030] ​Specifically, the actual driving environment data within the line-of-sight range is taken as input information of the corresponding preset machine learning model, and is substituted into the preset machine learning model to perform data fusion, environment feature fusion and environment feature extraction, so as to start training the preset machine learning model. After the training is completed, the preset driving environment prediction model is obtained. The preset driving environment prediction model can predict the driving environment within the beyond-line-of-sight range, and generate a virtual field of view within the beyond-line-of-sight range representing the predicted driving environment.

[0031] Before generating the virtual field of view within the beyond-line-of-sight range representing the predicted driving environment based on the preset driving environment prediction model and using the actual driving environment data within the line-of-sight range, the preset first deep learning model is also based on historical actual driving environment data within the line-of-sight range during vehicle driving in a historical period as input information of the preset first deep learning model, and historical actual driving environment data within the beyond-line-of-sight range is taken as output information of the preset first deep learning model, so as to obtain the preset driving environment prediction model through training of the preset first deep learning model.

[0032] In the embodiment of the present application, the corresponding preset machine learning model is the preset first deep learning model. The historical actual driving environment data within the line-of-sight range during vehicle driving in a historical period is taken as input information of the preset first deep learning model, and the historical actual driving environment data within the beyond-line-of-sight range is taken as output information of the preset first deep learning model, which is substituted into the preset first deep learning model to start training the preset first deep learning model. After the training is completed, the preset driving environment prediction model is obtained.

[0033] In the embodiment of the present application, the preset driving environment prediction model is composed of PointNet++ and a convolutional neural network framework.

[0034] In the embodiment of the present application, the virtual field of view is generated by using the following expression: ; Wherein, represents the virtual field of view, () represents an environment feature fusion and environment feature extraction function, represents point cloud data, represents image data, represents a timestamp synchronization factor, represents a coordinate transformation matrix.

[0035] Further, in step S130, the actual driving environment data within the line-of-sight range and the environment data belonging to the obstacle in the virtual field of view are marked, and then based on the preset obstacle detection model, the driving environment data carrying the mark is used to generate the corresponding real-time obstacle detection result.

[0036] Specifically, the embodiment first labels the actual driving environment data within the line-of-sight range and the environment data belonging to obstacles in the virtual field of view (for example: the position, speed, acceleration, etc. of the obstacles). Then, the labeled driving environment data is taken as the input information of the preset machine learning model, and is substituted into the preset machine learning model to perform environment feature fusion and environment feature extraction of the environment data belonging to obstacles, so as to start training the preset machine learning model, and after the training is completed, the preset obstacle detection model is obtained. The preset obstacle detection model can detect obstacles in the driving environment and generate corresponding real-time obstacle detection results.

[0037] Before generating the corresponding real-time obstacle detection results based on the preset obstacle detection model and using the labeled driving environment data, the application also labels the environment data belonging to obstacles in the historical actual driving environment data within the line-of-sight range and the historical actual driving environment data within the beyond-line-of-sight range, and then takes the labeled historical driving environment data as the input information of the preset second deep learning model, and takes the environment data belonging to obstacles as the output information of the preset second deep learning model, so as to obtain the preset obstacle detection model by training the preset second deep learning model.

[0038] In the embodiment of the application, the environment data belonging to obstacles in the historical actual driving environment data within the line-of-sight range and the historical actual driving environment data within the beyond-line-of-sight range is first labeled. At this time, the corresponding preset machine learning model is the preset second deep learning model. The labeled historical driving environment data within the line-of-sight range and the beyond-line-of-sight range is taken as the input information of the preset second deep learning model, and the environment data belonging to obstacles is taken as the output information of the preset second deep learning model, which is substituted into the preset second deep learning model to start training the preset second deep learning model, and after the training is completed, the preset obstacle detection model is obtained.

[0039] In the embodiment of the application, the preset second deep learning model adopts a Transformer variant (for example: Vision Transformer or Swin Transformer) framework based on a self-attention mechanism. The Transformer variant is used to realize the detection of obstacles, and the self-attention mechanism is used to capture the long-term dependence relationship in the obstacle trajectory, so as to further improve the detection accuracy of dynamic obstacles and realize the intention prediction and trajectory prediction of each obstacle.

[0040] Further, the types of driving environment data include but are not limited to point cloud data and video stream data, wherein, the application also adds YOLOv5, 3DSSD (3D Single Shot Detector) and DeepSORT (Deep Simple Online and Realtime Tracking) algorithms into the preset second deep learning model to obtain the spatial position change characteristics of each obstacle in the continuous period, and then realize the intention prediction and trajectory prediction of each obstacle.

[0041] YOLOv5 belongs to the YOLO series and is a real-time target detection algorithm based on convolutional neural network, which can simultaneously predict the position and category of multiple targets in one forward propagation. Its characteristics include fast speed, high accuracy, and is suitable for real-time video monitoring and automatic driving scenes. 3DSSD is a three-dimensional target detection algorithm specially designed for point cloud data, which directly operates on the original point cloud to generate three-dimensional bounding boxes for detecting targets such as pedestrians and vehicles. The algorithm design focuses on efficiency and accuracy, and has wide application in the fields of autonomous driving and robot navigation. DeepSORT is a multi-target tracking algorithm improved based on SORT, which combines Kalman filter and appearance features extracted by deep learning to stably and efficiently track multiple targets in video sequences, especially in handling occlusion and similar targets.

[0042] Therefore, by adding YOLOv5, 3DSSD and DeepSORT algorithms into the preset second deep learning model, the spatial characteristics of the detected obstacles are recognized and the detected obstacles are tracked, intention recognized and trajectory recognized, so as to ensure that the spatial position change characteristics of each obstacle in the continuous period are accurately and stably obtained in the complex traffic environment. According to the spatial position change characteristics of each obstacle in the continuous period, the intention prediction and trajectory prediction (such as moving direction, speed change, etc.) of each obstacle can be realized, which lays a good data foundation for early risk prediction and provides strong support for early warning decision.

[0043] In the embodiments of the application, the weight matrix is responsible for capturing long-term dependencies in the obstacle trajectory through self-attention mechanism, wherein the preset second deep learning model is represented by the following expression: ; wherein, represents the predicted position of the obstacle, represents the prediction function, represents the previous period obstacle trajectory data, represents the weight matrix.

[0044] Further, the point cloud data in the present application is collected by a vehicle-mounted laser radar arranged on the vehicle; and the video stream data is collected by a ground binocular camera arranged in a ground area along a running track of the vehicle.

[0045] In the embodiments of the present application, the vehicle-mounted laser radar arranged on the vehicle is used to collect the three-dimensional point cloud data in front of and around the vehicle in real time, so as to obtain the distance between different points in the driving environment and the vehicle; and the ground binocular camera arranged in the ground area along the running track of the vehicle is used to collect the video stream data in front of and around the vehicle in real time, so as to obtain the image data of the driving environment. The combination of the point cloud data and the video stream data can comprehensively and accurately identify the obstacles, greatly improving the range and accuracy of obstacle identification.

[0046] Further, the present application further predicts the collision time and the collision distance of the obstacle and the vehicle according to the generated real-time obstacle detection result; then, the first collision risk degree and the second collision risk degree are generated according to the collision time prediction result and the collision distance prediction result respectively, and then the comprehensive collision risk degree is obtained by analyzing and comparing each collision risk degree and a warning is given.

[0047] In the embodiments of the present application, the obstacle trajectory prediction result is extracted from the generated real-time obstacle detection result, and the collision time (Time to Collision, TTC) of the obstacle and the vehicle and the collision distance (Distance to Collision, DTC) of the obstacle and the vehicle are predicted respectively in combination with the vehicle speed. Then, the first collision risk degree is generated according to the collision time prediction result, and the second collision risk degree is generated according to the collision distance prediction result. Next, the collision risk degrees are analyzed and compared, the collision risk degree with higher risk degree is taken as the comprehensive collision risk degree, and a warning is given according to the preset mapping relationship between the collision time, the collision distance and the collision risk degree to prompt the current risk degree, so as to provide sufficient reaction time for the driver and effectively avoid collision accidents caused by sudden obstacles. For example, different levels of sound and light alarm signals are output by the vehicle-mounted warning instrument to prompt the current risk degree.

[0048] In one specific embodiment of the present application, the preset mapping relationship between the collision time, the collision distance and the collision risk degree is as follows: TTC greater than 10 seconds and DTC greater than 200 meters correspond to a low-risk collision risk degree; TTC between 5 seconds and 10 seconds and DTC between 100 meters and 200 meters correspond to a medium-risk collision risk degree; TTC less than or equal to 5 seconds and DTC less than or equal to 100 meters correspond to a high-risk collision risk degree.

[0049] Furthermore, this application also monitors key data affecting the reliability of early warning in real time based on a multi-head attention mechanism, and uses a parallel processing method for different types of key data to assess global risk. Then, it uses the global risk assessment results to verify the current risk level and issues an early warning based on the verified risk level.

[0050] In one specific embodiment of this application, based on a multi-head attention mechanism, at least four types of key data affecting the reliability of the warning are used: a first attention head monitors in real time a first parameter related to the quality and fusion effect of the fused data; a second attention head monitors in real time a second parameter related to changes in environmental data in the virtual field of view; a third attention head monitors in real time a third parameter related to the motion state of obstacles; and a fourth attention head monitors in real time a fourth parameter related to changes in the motion state of the vehicle. While obtaining the key data, a parallel processing approach is adopted for different types of key data, processing multiple data dimensions in parallel. This improves the processing response speed and accuracy while assessing the global risk. The global risk assessment result is then used to verify the current risk level. Specifically, based on a preset mapping relationship between the global risk assessment result and the collision risk level, a new risk level is generated. The rationality of the current risk level is verified by comparing its consistency with the new risk level. If the new risk level is consistent, a warning is issued directly based on the current risk level; otherwise, the new risk level is used as the verified risk level, and a warning is issued based on the new risk level. As can be seen, the multi-head attention mechanism enables this application to dynamically adapt to changes in the driving environment, ensuring the accuracy of early warning decisions in response to various emergencies in complex driving scenarios.

[0051] In this embodiment, the following expression is used to assess global risk by weighted aggregation of different types of key data: ; in, This indicates the overall risk assessment result. n Indicates the number of attention heads. Indicates the first i The weighting coefficient of each attention head. ( ) indicates the first ( ). i The output of each attention head, , , and These represent the first [item] in different types of key data. i Data in multiple dimensions.

[0052] In one specific embodiment of this application, different dimensions of information can be focused on independently, and based on the importance of each dimension, This is used to represent the degree of attention paid to information in a certain dimension. The first parameter, representing the first parameter related to the quality and fusion effect of the fused data, is the... i Data from multiple dimensions The second parameter, representing the change in environmental data within the virtual field of view, is the first... i Data from multiple dimensions The third parameter, which represents the motion state of the obstacle, is the first... i Data in each data dimension and The fourth parameter, which represents the change in the vehicle's motion state, is the first of its kind. i Data in multiple dimensions.

[0053] Example Two Based on the obstacle detection method for rail transit vehicles described in Embodiment 1 above, this application also provides an obstacle detection system for rail transit vehicles (hereinafter referred to as "obstacle detection system").

[0054] Figure 2 This is a block diagram of an obstacle detection system for rail transit vehicles according to an embodiment of this application. Figure 2 As shown, the obstacle detection system 20 in this embodiment includes: a data processing module 21, a driving environment prediction module 22, and an obstacle detection module 23. Specifically, the data processing module 21 is implemented according to the method described in step S110 above, and is configured to collect various types of driving environment data during vehicle operation in real time; the driving environment prediction module 22 is implemented according to the method described in step S120 above, and is configured to generate a virtual field of view representing the predicted driving environment beyond the line of sight using actual driving environment data within the line of sight range based on a preset driving environment prediction model; the obstacle detection module 23 is implemented according to the method described in step S130 above, and is configured to mark the actual driving environment data within the line of sight range and the environmental data belonging to obstacles in the virtual field of view, and then generate corresponding real-time obstacle detection results based on a preset obstacle detection model using the marked driving environment data.

[0055] The application discloses a kind of barrier detection method and system for rail transit vehicle.The method is based on vehicle-ground cooperation real-time collection vehicle driving process in a variety of types of driving environment data.And then, based on deep learning algorithm, using historical actual driving environment data and the environment data belonging to barrier in the process of vehicle driving in historical period, train out preset driving environment prediction model and preset barrier detection model, obtain corresponding real-time barrier detection result based on this.The application effectively improves the accuracy of dynamic barrier detection in the process of vehicle driving, can accurately identify barrier in the range of over-the-horizon in the process of rail transit vehicle operation, to reduce the accident caused by barrier, reduce the outage and maintenance cost caused by accident provide technical support.With the improvement of intelligent level of rail transit, the market demand of related equipment and technology will continue to grow, the application effectively improves the safety of rail transit system while improving transportation efficiency, will have wide application prospect in the future (especially in the research and development of intelligent rail transit system and automatic driving train), with higher market value.

[0056] The above is only the preferred embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the application, which should be covered within the protection scope of the application. Therefore, the protection scope of the application should be limited by the protection scope of the claims.

[0057] Of course, the application can have other various embodiments, without departing from the spirit and essence of the application, those skilled in the art can make various corresponding changes and modifications according to the application, but these corresponding changes and modifications should belong to the protection scope of the claims of the application.

[0058] Those skilled in the art should understand that the above-mentioned modules or steps of the application can be realized by general computing devices, which can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, alternatively, they can be realized by program code executable by computing device, so that they can be stored in storage device and executed by computing device, or they can be made into each integrated circuit module, or multiple modules or steps among them can be made into a single integrated circuit module to realize. Thus, the application is not limited to any specific hardware and software combination.

[0059] Although the present application has been described with reference to the above embodiments, the contents described above are only the embodiments adopted for facilitating the understanding of the present application, and are not intended to limit the present application. Any modification and change in the form and details of the embodiments can be made by any person skilled in the art without departing from the spirit and scope of the present application, and the patent protection scope of the present application shall be subject to the scope defined by the claims.

Claims

1. A method for obstacle detection for a rail vehicle, characterized in that, The method comprises: real-time acquisition of various types of driving environment data during vehicle driving; based on a preset driving environment prediction model, using the actual driving environment data within the visual range to generate a virtual field of view representing the predicted driving environment within the extra-visual range; labeling the actual driving environment data within the visual range and the environment data belonging to obstacles in the virtual field of view, and then based on a preset obstacle detection model, using the labeled driving environment data to generate corresponding real-time obstacle detection results.

2. The obstacle detection method according to claim 1, wherein Before generating a virtual field of view representing the predicted driving environment within the extra-visual range based on a preset driving environment prediction model using the actual driving environment data within the visual range, the obstacle detection method further comprises: based on a preset first deep learning model, using the historical actual driving environment data within the visual range during vehicle driving in the historical period as the input information of the preset first deep learning model, and using the historical actual driving environment data within the extra-visual range as the output information of the preset first deep learning model, so as to obtain the preset driving environment prediction model by training the preset first deep learning model; wherein the preset driving environment prediction model adopts PointNet++ and a convolutional neural network framework.

3. The obstacle detection method according to claim 2, wherein Before generating corresponding real-time obstacle detection results based on a preset obstacle detection model using the labeled driving environment data, the obstacle detection method further comprises: labeling the environment data belonging to obstacles in the historical actual driving environment data within the visual range and the historical actual driving environment data within the extra-visual range, and then based on a preset second deep learning model, using the labeled historical driving environment data as the input information of the preset second deep learning model, and using the environment data belonging to obstacles as the output information of the preset second deep learning model, so as to obtain the preset obstacle detection model by training the preset second deep learning model; wherein the preset second deep learning model adopts a Transformer variant framework based on a self-attention mechanism.

4. The obstacle detection method according to claim 3, wherein The types of driving environment data include but are not limited to point cloud data and video stream data, wherein the obstacle detection method further comprises: adding YOLOv5, 3DSSD and DeepSORT algorithms into the preset second deep learning model to obtain the spatial position change characteristics of each obstacle in the continuous period, and then realizing the intention prediction and trajectory prediction of each obstacle.

5. The obstacle detection method according to any one of claims 1 to 4, characterized in that, The obstacle detection method further comprises: predicting the collision time and collision distance of the obstacle and the vehicle according to the generated real-time obstacle detection results; generating a first collision risk degree and a second collision risk degree according to the collision time prediction result and the collision distance prediction result respectively, and then obtaining a comprehensive collision risk degree by analyzing and comparing each collision risk degree and giving a warning.

6. The obstacle detection method according to claim 5, wherein The obstacle detection method further comprises: The key data influencing the reliability of early warning is monitored in real time based on a multi-head attention mechanism, global risk is evaluated by using a parallel processing mode for different types of key data, the current risk degree is checked by using the global risk evaluation result, and early warning is performed according to the checked risk degree.

7. The obstacle detection method according to claim 6, wherein The global risk is evaluated by using the following expression: ； wherein, represents the global risk assessment result, n represents the number of attention heads, represents the weight coefficient of the i th attention head, represents the output of the i th attention head, , , and respectively represent the data of the i th data dimension in different types of key data.

8. The obstacle detection method according to any one of claims 1 to 4, characterized by, Before generating the virtual visual field representing the predicted driving environment in the beyond visual range, the obstacle detection method further comprises: Aligning all driving environment data in a unified space-time coordinate system to obtain the best virtual visual field.

9. The obstacle detection method according to claim 4, wherein, The obstacle detection method further comprises: The point cloud data is collected by a vehicle-mounted laser radar arranged on the vehicle; The video stream data is collected by a ground binocular camera arranged in a ground area along the running track of the vehicle.

10. An obstacle detection system for rail transit vehicles, characterized in that, The obstacle detection system comprises the following modules: A data processing module for collecting various types of driving environment data in real time during the driving of the vehicle; A driving environment prediction module for generating a virtual visual field representing the predicted driving environment in the beyond visual range based on a preset driving environment prediction model and using the actual driving environment data within the visual range; An obstacle detection module for marking the actual driving environment data within the visual range and the environment data belonging to obstacles in the virtual visual field, and then generating corresponding real-time obstacle detection results based on a preset obstacle detection model and using the driving environment data carrying the marks.

Images