Obstacle detection method and obstacle detection system for track

Abstract

The invention discloses an obstacle detection method and an obstacle detection system for a track, relates to the technical field of track detection, and is used for solving the problem of insufficient safety of track traffic. Comprising a data acquisition module, a data storage module, a data processing and fusion module, an intelligent analysis and decision module and an execution module which are in signal connection. The data acquisition module dynamically adjusts the working range and frequency of the sensor according to environmental changes. The collected data is processed and fused through the data processing and fusion module to generate a consistent orbit environment model, and a high-precision obstacle model is generated in combination with multi-source data. The intelligent analysis and decision-making module carries out classification, tracking and risk assessment on obstacles and generates a reasonable coping strategy according to the risk level. And the execution module executes related operation according to the generated strategy instruction. According to the method, track obstacles can be accurately detected, self-adaptive adjustment can be carried out according to dynamic environment changes, and the safety and operation efficiency of track traffic are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of track detection, and more specifically, to a method and a system for detecting obstacles on tracks. Background Art

[0002] With the rapid development of rail transit, the safety issue of the environment around the tracks has become increasingly prominent. Obstacle detection is crucial in the safety management of rail transit. Traditional obstacle detection technologies often rely on a single sensor, resulting in insufficient detection accuracy and reliability in complex environments. Existing detection methods such as vision and radar technologies can provide preliminary information about obstacles, but in bad weather, low light, or dynamic environments, false judgments or missed detections are likely to occur. In addition, existing technologies mostly adopt fixed sensor layouts and scanning strategies, which cannot flexibly adapt to environmental changes and result in significant resource waste. In order to improve the safety and efficiency of rail transit, there is an urgent need for an obstacle detection method that can perceive in real time, accurately locate, and adapt to dynamic environmental changes.

[0003] Therefore, an obstacle detection method for tracks based on multi-sensor fusion and dynamic adjustment has become an urgent need. By introducing an adaptive sensor layout, data cleaning, and spatio-temporal fusion technologies, the present invention aims to solve the deficiencies in the existing technologies, improve the detection accuracy and response speed of track obstacles, and ensure the safe operation of the rail transit system.

[0004] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the existing technologies, embodiments of the present invention provide a method and a system for detecting obstacles on tracks to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a preferred embodiment, it includes: a data acquisition module, a data processing and fusion module, an intelligent analysis and decision-making module, an execution module, and a data storage module, and the modules are signal-connected to each other;

[0008] The data storage module is used to store all data during the processing process;

[0009] The data acquisition module comprehensively perceives the track environment, uses a variety of sensors to collect information about the environment around the tracks, automatically adjusts the scanning range and scanning frequency of the sensors according to environmental changes and the dynamic behavior of obstacles, and collects obstacle data in real time to provide preliminary obstacle data;

[0010] The data processing and fusion module cleans and synchronizes the time of the obstacle data collected by the data acquisition module, and adopts a spatio-temporal adaptive fusion algorithm to dynamically adjust the data fusion strategy according to different spatio-temporal backgrounds, generating a highly consistent orbital environment model;

[0011] Based on the obstacle data after cleaning and fusion, the intelligent analysis and decision-making module identifies the type of obstacle, evaluates the risk level, and generates coping strategies;

[0012] The execution module quickly executes the response instructions through the instructions obtained by the intelligent analysis module.

[0013] In a preferred embodiment, the data acquisition module is based on TOF technology. The lidar rotates 360° around its own axis at a fixed frequency to capture the point cloud data of the surrounding environment of the track, and preliminarily aggregates the point cloud based on the DBSCAN clustering algorithm to generate the three-dimensional contour information of the candidate obstacle; a vision camera is used to collect the video images along the track, and a frame rate control mechanism is adopted to ensure data continuity and high resolution; at the same time, an infrared camera supplements the data in the night or low-light environment, and the data of the vision camera and the infrared camera are compared through the feature matching algorithm SIFT; a millimeter-wave radar is used to detect long-distance and high-speed moving targets through the Doppler effect, emits electromagnetic waves at a fixed frequency, receives the echo signal reflected by the target, performs Fourier transform on the echo signal, and extracts the speed, distance, and direction information of the target; ultrasonic sensors emit ultrasonic pulses at a specific frequency, receive the echoes reflected by the obstacles, calculate the distance of the obstacles according to the pulse propagation time, and generate the shape and azimuth information of the near-field obstacles through the data fusion of multiple ultrasonic sensors.

[0014] In a preferred embodiment, the data acquisition module collects the three-dimensional point cloud data around the track through the lidar, calculates the position x, y, z coordinates and size of the obstacle, and uses the DBSCAN algorithm to cluster the point cloud to identify the obstacle and determine the dense area; detects the speed and direction of the obstacle through the Doppler effect to identify objects approaching or moving away quickly; identifies static objects and dynamic behaviors in the environment through the YOLO algorithm of the target detection model to supplement the detailed features;

[0015] The data acquisition module sets a distance threshold Yj and a speed threshold Ys. When the distance of the obstacle is less than or equal to the distance threshold Yj, the scanning range is increased; when the distance of the obstacle is greater than the distance threshold Yj, the scanning range is reduced; when the speed of the obstacle is less than or equal to the speed threshold Ys, the data acquisition module reduces the scanning frequency through the Doppler effect, and when the speed of the obstacle is greater than the speed threshold Ys, the data acquisition module increases the scanning frequency through the Doppler effect;

[0016] Set the light intensity threshold Yg and the rainfall threshold Yr. When the ambient light intensity is less than or equal to the light intensity threshold Yg, the data acquisition module will enable the infrared camera; when the rainfall is greater than or equal to the rainfall threshold Yr, the data acquisition module will reduce the usage frequency of the vision sensor and rely more on the millimeter-wave radar for obstacle detection.

[0017] In a preferred embodiment, the data processing and fusion module first uses the RANSAC algorithm to assume that most points in the point cloud data belong to the same geometric model, and iteratively selects data subsets for fitting. Finally, the point set that conforms to the model is determined, and points that are isolated and do not match the main data set are identified as noise and removed from the data; for data from the vision camera, the data processing and fusion module uses the CLAHE adaptive histogram equalization technique to enhance the contrast of the image, making the image details clearer, and applies median filtering or Gaussian filtering in the image filtering algorithm to remove noise in the image. Then, by calculating the geometric shape of the point cloud, the boundaries of areas such as the ground and obstacles are identified to perform boundary detection on the point cloud data and optimize the point cloud data.

[0018] The data processing and fusion module uses the PTP precise time protocol to ensure the alignment of the timestamps of each sensor; at the same time, data interpolation processing is performed for sensors with different frequencies to fill the time gaps in the data of low-frequency sensors.

[0019] The data processing and fusion module fuses the position information of the lidar and the millimeter-wave radar through Kalman filtering, recursively estimates the state variables, eliminates measurement errors, and improves the positioning accuracy.

[0020] The data processing and fusion module uses the self-attention mechanism Self-Attention to fuse the image data and the lidar point cloud data, extracts the mutual relationships between sensors through the self-attention mechanism, and obtains the fused environmental model.

[0021] The data processing and fusion module is based on the feature learning of convolutional neural networks and deep learning algorithms, and automatically adjusts the weights of the information of each sensor through training the model.

[0022] In a preferred embodiment, the data processing and fusion module extracts the spatio-temporal features of the data collected by each sensor, including timestamps, spatial position information, and environmental conditions, and analyzes the changes in spatio-temporal features according to different environmental conditions and spatio-temporal backgrounds.

[0023] The data processing and fusion module determines the weight value of each sensor under specific spatio-temporal conditions according to time, space, environment and speed factors. Specifically, it is based on the formula: Q = α×W1 + β×W2 + γ×W3 + δ×W4, where W represents the weight value of the sensor, α, β, γ, δ represent adjustment coefficients, W1 represents the weight of the time factor; W2 represents the weight of the space factor; W3 represents the weight of the obstacle speed factor, and W4 represents the weight of the environment factor;

[0024] The data processing and fusion module first needs to synchronize the sensor data in time through linear interpolation or nearest neighbor interpolation methods, and then use the spatial position and attitude information of the sensors to perform coordinate transformation to convert the data of all sensors into the same global coordinate system.

[0025] The data processing and fusion module obtains a comprehensive result Dfused by means of weighted average based on the weight value of the sensor under specific spatio-temporal conditions calculated and the corresponding perceptual data collected by the sensor. Specifically, it is based on the formula where Di represents the perceptual data of the sensor.

[0026] In a preferred embodiment, the intelligent analysis and decision-making module assigns a unique ID to each detected obstacle and records the initial state: position, speed, direction; then, it uses the DeepSORT algorithm to perform multi-frame tracking on the obstacle, combines deep learning to extract features for target re-identification, and uses Kalman filtering for position estimation to generate the trajectory data of the target;

[0027] Based on the position and speed information of the obstacle obtained by tracking, the intelligent analysis and decision-making module calculates the position and speed of the obstacle at several future moments through a uniform or accelerated motion model.

[0028] The intelligent analysis and decision-making module determines the risk value by calculating the position and speed of the obstacle through weighted average. Specifically, it is based on the formula: risk value = distance weight × (1 / distance) + speed weight × speed, and sets a risk threshold Yfx. When the risk value is less than or equal to the risk threshold Yfx, the risk level is low risk at this time. When the risk value is greater than the risk threshold Yfx, the risk level is high risk at this time;

[0029] The intelligent analysis and decision-making module generates reasonable corresponding policy instructions according to the risk level, and transmits the generated policy instructions to the execution module and monitors its execution situation.

[0030] In a preferred embodiment, the execution module first parses the communication data packet according to the output instruction of the intelligent analysis module using the protocol standard CAN bus protocol, extracts the instruction type and parameters. When executing a high-risk level instruction, the execution module calculates the maximum braking force based on the current speed and load condition of the train; calls the train braking controller to trigger the emergency braking mode; cooperates with the train anti-lock braking mechanism to adjust the braking force in real time to prevent wheel skidding; when executing a low-risk level instruction, the execution module stores the obstacle information in the train recorder, and the train maintains its current operating state.

[0031] The execution module collects the train status data in real time through speed sensors, braking state sensors, and position sensors; sends the monitoring data to the TCMS via the CAN bus, and ensures that the status signal is returned after the instruction execution is completed.

[0032] In a preferred embodiment, Step 1: Dynamically obtain the track environment information through multiple sensors, and adjust the acquisition strategy according to the environmental conditions;

[0033] Step 2: Clean the noise of the sensing data, unify the time and space coordinates, and ensure data consistency;

[0034] Step 3: Use the weighted fusion and deep learning methods to dynamically integrate multi-source data and generate an accurate environment model;

[0035] Step 4: Detect and classify obstacles, evaluate the risk level, and generate countermeasures such as alarm or braking instructions.

[0036] The present invention discloses an obstacle detection method and an obstacle detection system for tracks, which relates to the technical field of environmental detection and is used to solve the problem of insufficient safety of rail transit; it includes: a data acquisition module, a data storage module, a data processing and fusion module, an intelligent analysis and decision-making module, and an execution module, and the signals between the modules are connected. The data acquisition module dynamically adjusts the working range and frequency of the sensors according to the environmental changes. The collected data is used by the data processing and fusion module to generate a consistent track environment model, and a high-precision obstacle model is generated by combining multi-source data. The intelligent analysis and decision-making module classifies, tracks, and evaluates the risks of obstacles, and generates reasonable countermeasures according to the risk level. The execution module performs relevant operations according to the generated strategy instructions. This method can accurately detect track obstacles and make adaptive adjustments according to the dynamic environmental changes, significantly improving the safety and operation efficiency of rail transit. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic structural diagram of the obstacle detection method and the obstacle detection system for tracks of the present invention.

[0038] Figure 2This is the flowchart of the obstacle detection method and the operation of the obstacle detection method for the track of the present invention. Detailed implementation mode

[0039] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0040] Embodiment

[0041] The present invention discloses an obstacle detection system for tracks, including: a data acquisition module, a data processing and fusion module, an intelligent analysis and decision-making module, an execution module, and a data storage module, and the modules are connected by signals.

[0042] The data storage module is used to store all data during the processing process.

[0043] The data acquisition module comprehensively senses the track environment, uses a variety of sensors to collect information on the surrounding environment of the track, collects obstacle data in real time, and provides preliminary obstacle data.

[0044] Specifically, in the data acquisition module, first, based on TOF (Time of Flight) technology, the lidar rotates 360° around its own axis at a fixed frequency (such as 10Hz) to capture the point cloud data of the surrounding environment of the track, and based on an algorithm (such as DBSCAN clustering), the point cloud is preliminarily aggregated to generate the three-dimensional contour information of the candidate obstacle; further, a vision camera is used to collect the video images along the track, and a frame rate control mechanism (such as 30FPS) is adopted to ensure data continuity and high resolution. At the same time, an infrared camera supplements the data in the night or low-light environment to ensure all-day perception ability. Through a feature matching algorithm (such as SIFT or ORB), the data of the vision camera and the infrared camera are compared to enhance the detection accuracy; further, a millimeter-wave radar is used to detect long-distance and high-speed moving targets through the Doppler effect, emits electromagnetic waves at a fixed frequency, receives the echo signal reflected by the target, performs a Fourier transform on the echo signal, and extracts the speed, distance, and direction information of the target; ultrasonic sensors emit ultrasonic pulses at a specific frequency, receive the echo reflected by the obstacle, calculate the distance of the obstacle according to the pulse propagation time, and generate the shape and orientation information of the near-field obstacle through the data fusion of multiple ultrasonic sensors.

[0045] It should be noted that the sensor layout is usually fixed, which may lead to detection blind spots in areas with dense obstacles or when the environment changes significantly. At the same time, the fixed scanning frequency and scanning range will also cause excessive waste of resources during the process of detecting obstacles. In this embodiment, a dynamic adaptive sensor layout mechanism is designed, and an adjustable sensor array (such as lidar, camera, and millimeter-wave radar) is installed. The data acquisition module automatically adjusts the scanning range and scanning frequency of the sensors according to environmental changes and the dynamic behavior of obstacles to ensure the best data acquisition effect at all times.

[0046] First, the data acquisition module monitors the track environment in real time, including the position, type, speed of obstacles, and environmental factors (such as light, weather, etc.), and calculates the optimal sensor layout based on the real-time data.

[0047] Specifically, the data acquisition module collects three-dimensional point cloud data around the track through lidar, calculates the position (x, y, z coordinates) and size of the obstacles, and uses the DBSCAN algorithm to cluster the point cloud to identify obstacles and determine the dense areas; detects the speed and direction of the obstacles through the Doppler effect to identify objects approaching or moving away quickly; uses a target detection model (such as the YOLO algorithm) to identify static objects and dynamic behaviors in the environment and supplement detailed features. For example, during the day, it can identify stationary objects such as trees and luggage, and at night, it can detect heat sources through infrared imaging; uses meteorological sensors and optoelectronic sensors to detect the temperature, wind speed, and light intensity in the environment. For example: there are two obstacles within 50 meters ahead, with sizes of (1.5m×0.8m×0.6m) and (2.0m×0.7m×0.5m) respectively, the speed of target 1 is 20 km / h; the speed of target 2 is 15 km / h, target 1 is an animal (deer), target 2 is a stationary rock, the meteorological sensor detects the current wind speed is 3 m / s, and the rainfall is 6 mm / h. The optoelectronic sensor detects the current light intensity is 8 lux (low light).

[0048] Furthermore, based on the acquired environmental change data and obstacle dynamic behavior data, the data acquisition module dynamically adjusts the scanning range and working angle, sets the distance threshold Yj and speed threshold Ys. When the distance of the obstacle is less than or equal to the distance threshold Yj, the scanning range is increased to ensure detection accuracy and improve the ability to identify obstacles. When the distance of the obstacle is greater than the distance threshold Yj, the scanning range is reduced to avoid collecting excessive irrelevant data, thereby saving computing resources and improving efficiency. When the speed of the obstacle is less than or equal to the speed threshold Ys, the data acquisition module reduces the scanning frequency through the Doppler effect to avoid wasting resources and improve the overall processing efficiency. When the speed of the obstacle is greater than the speed threshold Ys, the data acquisition module increases the scanning frequency through the Doppler effect to ensure timely detection of potential dangerous objects. For example, in this embodiment, the distance threshold Yj is set to 30, and the speed threshold Ys is set to 20 km / h. By comparing the acquired sample data with the thresholds, it is obtained that the distance of target 1 is 40 meters, which is greater than the distance threshold, and the speed is 20 km / h, which is equal to the speed threshold. At this time, the data acquisition module automatically reduces the lidar scanning range to ±45° and reduces the scanning frequency of the millimeter-wave radar through the Doppler effect.

[0049] Meanwhile, under specific environmental conditions, certain sensors are automatically enabled or disabled. Specifically, in this embodiment, the light intensity threshold Yg and rainfall threshold Yr are set. When the environmental light intensity is less than or equal to the light intensity threshold Yg, the data acquisition module enables the infrared camera to obtain supplementary data in the night or low-light environment and improve the environmental perception ability. When the rainfall is greater than or equal to the rainfall threshold Yr, the data acquisition module reduces the usage frequency of the vision sensor and relies more on the millimeter-wave radar for obstacle detection.

[0050] The data processing and fusion module cleans, synchronizes in time, and fuses the obstacle data collected by the data acquisition module to generate a highly consistent orbital environment model.

[0051] Specifically, the data processing and fusion module first uses the RANSAC algorithm to assume that most points in the point cloud data belong to the same geometric model (such as a plane or a curve), and iteratively selects data subsets for fitting. Finally, the point set that conforms to the model is determined, and the isolated points that do not match the main data set are identified as noise and removed from the data. Furthermore, for the data from the vision camera, the CLAHE (Contrast Limited Adaptive Histogram Equalization) technique is used to enhance the contrast of the image, making the image details clearer, especially in low-light or high-contrast scenes, and an image filtering algorithm (such as median filtering or Gaussian filtering) is applied to further remove the noise in the image and maintain the main features of the image. Then, by calculating the geometric shape of the point cloud, the boundaries of areas such as the ground and obstacles are identified to perform boundary detection on the point cloud data and optimize the point cloud data for subsequent obstacle detection and classification.

[0052] Furthermore, the data processing and fusion module uses PTP (Precision Time Protocol) to ensure the alignment of timestamps of each sensor, guarantee the synchronization of multi-sensor data, and enable the data collected by multiple sensors to be processed at the same time point. Meanwhile, data interpolation processing is performed on sensors with different frequencies (for example, the sampling frequencies of lidar data and visual data are different) to fill the time gaps in the data of low-frequency sensors. For example, if the lidar samples at a frequency of 10 Hz and the camera samples at a frequency of 5 Hz, after data interpolation processing, the camera data can be evenly inserted at time intervals, so that all data can be matched and fused.

[0053] After completing data cleaning, the data processing and fusion module fuses the position information of the lidar and millimeter-wave radar through Kalman filtering, recursively estimates state variables (such as position, speed, etc.), eliminates measurement errors, and improves positioning accuracy. For example, assume that the position of an obstacle measured by the lidar is (x1, y1), and the position of the obstacle measured by the millimeter-wave radar is (x2, y2). Through Kalman filtering, the data processing and fusion module will perform weighted fusion on these two measurement values according to their measurement errors to obtain a more accurate position of the obstacle.

[0054] Furthermore, the data processing and fusion module uses the self-attention mechanism (Self-Attention) to fuse image data and lidar point cloud data, extracts the mutual relationships between sensors through the self-attention mechanism, and obtains a fused environmental model.

[0055] After completing data fusion, the data processing and fusion module is based on the feature learning of convolutional neural networks and deep learning algorithms. By training the model, it can find the rules between different perception data, automatically adjust the weights of each sensor's information, and achieve efficient and accurate environmental modeling during the fusion process.

[0056] Finally, the data processing and fusion module outputs a complete obstacle model.

[0057] It should be noted that a fixed fusion strategy is likely to lead to perception failure or increased errors in dynamic and complex environments. In this embodiment, according to the acquisition time, spatial position, and environmental conditions of different sensors, a spatio-temporal adaptive fusion algorithm is adopted to dynamically adjust the data fusion strategy according to different spatio-temporal backgrounds (such as day and night, open and crowded environments). This ensures high-precision perception and dynamic tracking of obstacles in different environments.

[0058] First, the data processing and fusion module extracts the spatio-temporal features of the data collected by each sensor, including timestamps, spatial position information (the relative position between the sensor and the obstacle), and environmental conditions (light, rainfall, etc.).

[0059] Specifically, since light changes can affect the recognition ability of the camera, weather changes may affect the accuracy of lidar, and changes in the distance of moving targets can affect the weights of sensors. Therefore, the data processing and fusion module analyzes the changes in spatio-temporal features according to different environmental conditions and spatio-temporal backgrounds. For example:

[0060] Day and night have different effects on different sensors: During the day, due to good lighting, the weight of the visual camera is relatively high; at night, the performance of the visual camera will be greatly reduced, so the weights of infrared sensors and millimeter-wave radars can be increased.

[0061] When the obstacle is close to the sensor, the data of sensors close to the obstacle (such as lidar and visual cameras) are used, which can provide higher-precision perception data; when the obstacle is far from the sensor, millimeter-wave radars or other long-range sensors are used, which are more suitable for long-range perception.

[0062] Weather conditions such as rain, snow, and haze will have a greater impact on the perception effect of the visual camera. Therefore, the weight of the visual camera is reduced, and the weights of millimeter-wave radars or infrared sensors are increased. During the day with good lighting conditions, the visual sensor performs better; at night, the infrared sensor performs better. Therefore, the weights need to be adjusted according to light changes. At the same time, when the obstacle moves at a high speed, the weight of the millimeter-wave radar is relatively high because it has stronger adaptability to fast-moving targets.

[0063] Furthermore, the data processing and fusion module determines the weight values of each sensor under specific spatio-temporal conditions according to time, space, environment, and speed factors. Specifically, based on the formula: Q = α×W1 + β×W2 + γ×W3 + δ×W4, where W represents the weight value of the sensor, α, β, γ, δ represent adjustment coefficients, W1 represents the weight of the time factor, which depends on whether it is day or night; W2 represents the weight of the space factor, which is based on the distance between the sensor and the obstacle; W3 represents the weight of the obstacle speed factor, and when the obstacle speed is high, the radar weight is increased; W4 represents the weight of the environmental factor.

[0064] Furthermore, the first step in spatio-temporal data fusion is to ensure that the data from different sensors are aligned in time and space, so that the data can be compared and fused in the same spatio-temporal coordinate system. Therefore, the data processing and fusion module first needs to synchronize the sensor data in time through linear interpolation or nearest-neighbor interpolation methods, and then use the spatial position and attitude information of the sensors to perform coordinate transformation to convert the data of all sensors into the same global coordinate system. It should be noted that when the visual camera and lidar are on the same platform, the attitude information of the platform (such as azimuth angle and position) needs to be used to convert their data into the global coordinate system.

[0065] Further, after the spatio-temporal feature synchronization, the data processing and fusion module obtains a comprehensive perception result Dfused through weighted averaging of the weight values of the sensors under specific spatio-temporal conditions calculated and the corresponding perception data collected by the sensors, specifically according to the formula where Di represents the perception data of the sensor.

[0066] The intelligent analysis and decision-making module identifies the obstacle type, evaluates the risk level, and generates countermeasures based on the obstacle data after cleaning and fusion;

[0067] Since YOLOv5 has a good balance between real-time performance and detection accuracy and is suitable for fast classification in dynamic environments, the intelligent analysis and decision-making module first preprocesses the image data provided by the sensors (such as normalization, size adjustment), and inputs the preprocessed data into the deep learning model YOLOv5 to output the category and bounding box information of the obstacles, obtaining the obstacle markings and their spatial positions of categories such as pedestrians, vehicles, and sundries.

[0068] Further, the intelligent analysis and decision-making module dynamically tracks the identified obstacles through the DeepSORT algorithm and predicts their future trajectories:

[0069] First, the intelligent analysis and decision-making module assigns a unique ID to each detected obstacle and records the initial state (position, speed, direction). Then, the DeepSORT algorithm is used to perform multi-frame tracking on the obstacles, combined with deep learning to extract features for target re-identification to ensure the continuity of tracking. Kalman filtering is used for position estimation to generate the trajectory data of the target.

[0070] Further, the intelligent analysis and decision-making module calculates the positions and speeds of the obstacles at several future moments through a uniform or acceleration motion model based on the obstacle position and speed information obtained from tracking.

[0071] Further, the intelligent analysis and decision-making module evaluates the risk level by combining the obstacle position and speed information.

[0072] Specifically, the intelligent analysis and decision-making module calculates the risk value by weighted averaging the position and speed of the obstacle, specifically according to the formula: risk value = distance weight × (1 / distance) + speed weight × speed, and sets a risk threshold Yfx. When the risk value is less than or equal to the risk threshold Yfx, the risk level is low risk, and the obstacle is far from the track or has no dangerous behavior; when the risk value is greater than the risk threshold Yfx, the risk level is high risk, and at this time the obstacle is close to the center line of the track and has a high speed or a direction directly pointing to the track.

[0073] Finally, the intelligent analysis and decision-making module generates reasonable corresponding policy instructions according to the risk level, and transmits the generated policy instructions to the execution module and monitors its execution status.

[0074] Examples of policy instructions are as follows:

[0075] High risk: Trigger an emergency stop, activate the automatic braking system to ensure the safety of the train.

[0076] Low risk: Record the event in the log and maintain normal operation.

[0077] It should be noted that the hierarchical policy can ensure flexible response to different risk scenarios and avoid over-response.

[0078] The execution module quickly executes the response instructions based on the obtained instructions.

[0079] Specifically, the execution module first parses the communication data packet using a protocol standard (such as the CAN bus protocol) according to the output instructions of the intelligent analysis module, extracts the instruction type and parameters (speed adjustment value, braking force, alarm status, etc.). When executing a high-risk level instruction, the execution module calculates the maximum braking force based on the current speed and load of the train. It calls the train braking control system (Brake Control System, BCS) to trigger the emergency braking mode. Cooperating with the train anti-lock braking system (Anti-lock Braking System, ABS), it adjusts the braking force in real time to prevent the wheels from slipping; when executing a low-risk level instruction, the execution module stores the obstacle information (position, speed, risk level, etc.) in the train recording system (Event Data Recorder, EDR), and the train maintains its current operating state.

[0080] Furthermore, the execution module collects the train status data in real time through speed sensors, braking state sensors, and position sensors. It sends the monitoring data to the TCMS via the CAN bus to ensure that a status signal (such as braking completion, deceleration completion) is returned after the instruction execution is completed.

[0081] The present invention also relates to an obstacle detection method for railways. It includes:

[0082] Step 1: Dynamically obtain the track environment information through multiple sensors and adjust the acquisition strategy according to the environmental conditions.

[0083] Step 2: Clean the noise of the sensing data, unify the time and space coordinates to ensure data consistency.

[0084] Step 3: Use the weighted fusion and deep learning methods to dynamically integrate multi-source data and generate an accurate environmental model.

[0085] Step 4: Detect and classify obstacles, evaluate the risk level, and generate countermeasures such as alarm or braking instructions.

[0086] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0087] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product.

[0088] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and the inventive constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0089] In addition, the functional modules in each embodiment of this application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0090] As mentioned above, the above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application, and all should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0091] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. Obstacle detection system for tracks, characterized in that, Including: A data acquisition module, a data processing and fusion module, an intelligent analysis and decision-making module, an execution module, and a data storage module, with signal connections between the modules; The data storage module is used to store all data during the processing; The data acquisition module comprehensively senses the track environment, uses a variety of sensors to collect information on the surrounding environment of the track, automatically adjusts the scanning range and scanning frequency of the sensors according to environmental changes and the dynamic behavior of obstacles, and collects obstacle data in real time to provide preliminary obstacle data; The data processing and fusion module cleans and synchronizes the time of the obstacle data collected by the data acquisition module, and adopts a spatio-temporal adaptive fusion algorithm to dynamically adjust the data fusion strategy according to different spatio-temporal backgrounds to generate a highly consistent track environment model; The intelligent analysis and decision-making module identifies the type of obstacle, evaluates the risk level, and generates coping strategies based on the obstacle data after cleaning and fusion; The execution module quickly executes the coping instructions through the instructions obtained by the intelligent analysis module.

2. The obstacle detection system for a track according to claim 1, characterized in that: The data acquisition module is based on TOF technology. The lidar rotates 360° around its own axis at a fixed frequency to capture the point cloud data of the surrounding environment of the track, and preliminarily aggregates the point cloud based on the DBSCAN clustering algorithm to generate the three-dimensional contour information of the candidate obstacle; a visual camera is used to collect video images along the track, and a frame rate control mechanism is adopted to ensure data continuity and high resolution; at the same time, an infrared camera supplements data in night or low-light environments, and the data of the visual camera and the infrared camera are compared through the feature matching algorithm SIFT; a millimeter-wave radar detects long-distance high-speed moving targets through the Doppler effect, emits electromagnetic waves at a fixed frequency, receives the echo signal reflected by the target, performs Fourier transform on the echo signal, and extracts the speed, distance, and direction information of the target; ultrasonic sensors emit ultrasonic pulses at a specific frequency, receive the echoes reflected by obstacles, calculate the distance of the obstacles according to the pulse propagation time, and generate the shape and azimuth information of the near-field obstacles through the data fusion of multiple ultrasonic sensors.

3. The obstacle detection system for a track according to claim 2, characterized in that: The data acquisition module collects the three-dimensional point cloud data around the track through the lidar, calculates the position x, y, z coordinates and size of the obstacle, and uses the DBSCAN algorithm to cluster the point cloud to identify the obstacle and determine the dense area; detects the speed and direction of the obstacle through the Doppler effect to identify objects approaching or moving away quickly; Identifies static objects and dynamic behaviors in the environment through the YOLO algorithm of the target detection model to supplement detailed features; The data acquisition module sets a distance threshold Yj and a speed threshold Ys. When the distance of the obstacle is less than or equal to the distance threshold Yj, the scanning range is increased; when the distance of the obstacle is greater than the distance threshold Yj, the scanning range is reduced; when the speed of the obstacle is less than or equal to the speed threshold Ys, the data acquisition module reduces the scanning frequency through the Doppler effect, and when the speed of the obstacle is greater than the speed threshold Ys, the data acquisition module increases the scanning frequency through the Doppler effect; Set the light intensity threshold Yg and the rainfall threshold Yr. When the ambient light intensity is less than or equal to the light intensity threshold Yg, the data acquisition module will enable the infrared camera; when the rainfall is greater than or equal to the rainfall threshold Yr, the data acquisition module will reduce the usage frequency of the vision sensor and rely more on the millimeter-wave radar for obstacle detection.

4. The obstacle detection system for a track according to claim 3, wherein; The data processing and fusion module first uses the RANSAC algorithm to assume that most points in the point cloud data belong to the same geometric model, and iteratively selects data subsets for fitting. Finally, the point set that conforms to the model is determined. Points that are isolated and do not match the main data set are identified as noise and removed from the data; for data from the vision camera, the data processing and fusion module uses the CLAHE adaptive histogram equalization technique to enhance the contrast of the image, making the image details clearer, and applies median filtering or Gaussian filtering in the image filtering algorithm to remove the noise in the image. Then, by calculating the geometric shape of the point cloud, the boundaries of areas such as the ground and obstacles are identified to perform boundary detection on the point cloud data and optimize the point cloud data; The data processing and fusion module uses the PTP precise time protocol to ensure the alignment of the timestamps of each sensor; at the same time, data interpolation processing is performed on sensors with different frequencies to fill the time gaps in the data of low-frequency sensors; The data processing and fusion module fuses the position information of the lidar and the millimeter-wave radar through Kalman filtering, recursively estimates the state variables, eliminates measurement errors, and improves the positioning accuracy; The data processing and fusion module uses the self-attention mechanism Self-Attention to fuse the image data and the lidar point cloud data, extracts the mutual relationships between sensors through the self-attention mechanism, and obtains the fused environment model; The data processing and fusion module is based on the feature learning of the convolutional neural network and the deep learning algorithm, and automatically adjusts the weights of the information of each sensor through training the model.

5. The obstacle detection system for a track according to claim 4, characterized in that: The data processing and fusion module extracts the spatio-temporal features of the data collected by each sensor, including timestamps, spatial position information, and environmental conditions, and analyzes the changes in the spatio-temporal features according to different environmental conditions and spatio-temporal backgrounds; The data processing and fusion module determines the weight value of each sensor under specific spatio-temporal conditions according to time, space, environment, and speed factors. Specifically, according to the formula: Q = α×W1 + β×W2 + γ×W3 + δ×W4, where W represents the weight value of the sensor, α, β, γ, δ represent adjustment coefficients, W1 represents the weight of the time factor; W2 represents the weight of the space factor; W3 represents the weight of the obstacle speed factor, and W4 represents the weight of the environment factor; The data processing and fusion module first needs to synchronize the time of the sensor data by means of linear interpolation or nearest neighbor interpolation, and then uses the spatial position and attitude information of the sensor for coordinate transformation to convert the data of all sensors into the same global coordinate system. The data processing and fusion module obtains a comprehensive result Dfused by calculating the weight values of sensors under specific spatio-temporal conditions and the corresponding sensed data collected by the sensors through weighted averaging, specifically according to the formula where Di represents the sensed data of the sensor.

6. The obstacle detection system for a track according to claim 5, characterized in that: The intelligent analysis and decision-making module assigns a unique ID to each detected obstacle and records the initial state: position, speed, and direction. Then, it uses the DeepSORT algorithm to track the obstacle over multiple frames, extracts features using deep learning for target re-identification, and estimates the position using Kalman filtering to generate the trajectory data of the target. Based on the position and speed information of the obstacle obtained from tracking, the intelligent analysis and decision-making module calculates the position and speed of the obstacle at several future moments through a uniform or acceleration motion model. The intelligent analysis and decision-making module determines the risk value by calculating the position and speed of the obstacle through weighted averaging. Specifically, according to the formula: Risk value = distance weight × (1 / distance) + speed weight × speed, and sets the risk threshold Yfx. When the risk value is less than or equal to the risk threshold Yfx, the risk level is low risk at this time. When the risk value is greater than the risk threshold Yfx, the risk level is high risk at this time. The intelligent analysis and decision-making module generates reasonable corresponding policy instructions according to the risk level, and transmits the generated policy instructions to the execution module and monitors its execution status.

7. The obstacle detection system for a track according to claim 6, wherein: The execution module first parses the communication data packet using the protocol standard CAN bus protocol according to the output instructions of the intelligent analysis module, extracts the instruction type and parameters. When executing high-risk level instructions, the execution module calculates the maximum braking force according to the current speed and load condition of the train; calls the train braking controller to trigger the emergency braking mode. Cooperates with the train anti-lock braking mechanism to adjust the braking force in real time to prevent wheel skidding. When executing low-risk level instructions, the execution module stores the obstacle information in the train recorder, and the train maintains its current operating state. The execution module collects the train status data in real time through speed sensors, braking state sensors, and position sensors; sends the monitoring data to the TCMS through the CAN bus to ensure that the status signal is returned after the instruction execution is completed.

8. An obstacle detection method for tracks, characterized in that: Step 1: Dynamically obtain the track environment information through multiple sensors and adjust the acquisition strategy according to the environmental conditions. Step 2: Clean the noise of the sensing data, unify the time and space coordinates to ensure data consistency. Step 3: Use weighted fusion and deep learning methods to dynamically integrate multi-source data to generate an accurate environmental model. Step 4: Detect and classify obstacles, evaluate the risk level, and generate countermeasures such as alarm or braking instructions.

Images