A track inspection device autonomously plans a path

Abstract

This invention discloses an autonomous path planning track inspection device, comprising a mobile platform, a sensing module, a control and computing center, and a communication module. The sensing module includes a lidar and / or visual odometry for positioning and mapping, and a millimeter-wave radar, ultrasonic sensor, and visual camera for obstacle detection. The control and computing center includes a global path planner for generating a global reference path, a local path planner for real-time trajectory planning based on a local environment map fused from multiple sensors, and an emergency obstacle avoidance device for dynamic obstacle avoidance. This invention achieves fully autonomous operation in both known and unknown environments through a three-layer decision-making architecture of "global planning - local replanning - emergency response," significantly reducing reliance on manual remote control or fixed procedures.

Description

Technical Field

[0001] This invention belongs to the field of rail transit inspection technology and relates to a rail inspection device that autonomously plans its path. Background Technology

[0002] Traditional track inspection mainly relies on manual labor or inspection vehicles operating on fixed routes. Manual inspection is inefficient, unsafe, and highly susceptible to weather and lighting conditions. Existing automated inspection devices mostly rely on pre-programmed fixed routes or simple track following, lacking the ability to adapt to dynamic environmental changes. When temporary obstacles (such as fallen rocks or intrusions) appear on the track, or when the track itself experiences localized deformation or breakage, existing devices often fail to identify and respond effectively, potentially leading to equipment damage or interruption of the inspection mission. Therefore, there is an urgent need for an autonomous inspection device capable of intelligently planning paths, real-time obstacle avoidance, and adapting to complex track environments. Summary of the Invention

[0003] In order to overcome at least one deficiency of the prior art, the present invention provides a track inspection device that autonomously plans its path.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a track inspection device for autonomous path planning, comprising a mobile platform, a sensing module, a control and computing center, and a communication module; the sensing module includes a lidar and / or visual odometry for positioning and mapping, and a millimeter-wave radar, an ultrasonic sensor, and a visual camera for obstacle detection; the control and computing center includes a global path planner for generating a global reference path, a local path planner for real-time trajectory planning based on a local environment map fused from multiple sensors, and an emergency obstacle avoidance device for dynamic obstacle avoidance.

[0005] Furthermore, the local path planning layer adopts a dynamic window method to generate a smooth local trajectory while satisfying the dynamic constraints of the mobile platform and avoiding obstacles.

[0006] Furthermore, when the ultrasonic sensor detects a sudden nearby obstacle, the emergency obstacle avoidance device sends a braking command with the highest priority to the mobile platform.

[0007] Furthermore, the global path planner uses a search algorithm to perform global path planning based on the task start point, end point, and track network map, generating a global reference path; the local path planner uses multi-sensor fusion to perceive the real-time environment, constructs and updates a locally occupied grid map, and uses the global reference path as a guide to generate a collision-free local trajectory on the local map based on the dynamic constraints of the mobile platform.

[0008] Furthermore, the local path planner employs a dynamic window method, and its evaluation function G(v, ω) includes: Heading(v, ω), Dist(v, ω), Velocity(v, ω), and Semantic(v, ω); wherein the Semantic(v, ω) is calculated based on the classification results of obstacles by the visual sensor and the motion prediction of dynamic obstacles.

[0009] Furthermore, the emergency obstacle avoidance device uses linear extrapolation or Kalman filtering to predict the motion of dynamic obstacles.

[0010] Furthermore, the classification results of the obstacles generate different penalty weights. For dynamic obstacles, their future positions are linearly extrapolated or predicted by Kalman filtering. When evaluating the trajectory, the current obstacle position is checked, as well as its predicted position. For small, static obstacles that can be overcome, their penalty weights are reduced after visually identifying their category and size, allowing the device to approach or perform specific obstacle-crossing actions at extremely low speeds. For critical static obstacles, extremely high penalty weights are assigned to ensure that the planned path must stay away from them.

[0011] In summary, the advantages of this invention are: This invention achieves fully autonomous operation in both known and unknown environments through a three-tier decision-making architecture of "global planning - local replanning - emergency response," significantly reducing reliance on manual remote control or fixed procedures.

[0012] The multi-sensor fusion scheme of this invention ensures reliable sensing under different weather, lighting, and track environments (open-air, tunnel, bridge). Local real-time planning enables the device to flexibly respond to various static and dynamic obstacles that appear on the track.

[0013] This invention uses global planning to ensure overall task optimization and local planning to guarantee smooth and safe driving. Combined with a high-efficiency obstacle-crossing chassis, it can continuously complete long-distance inspection tasks and minimize the need for manual inspection in dangerous areas.

[0014] This invention integrates path navigation and track status detection functions, enabling "inspection while moving" and completing infrastructure inspection tasks in an integrated manner. Attached Figure Description

[0015] Figure 1 This is a modular schematic diagram of the device of the present invention. Detailed Implementation

[0016] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0017] Example: like Figure 1 As shown, a track inspection device that autonomously plans its path includes a mobile platform, a sensing module, a control and computing center, and a communication module. Mobile platforms are used for speed and position control.

[0018] The sensing module integrates multiple sensors for environmental perception and self-state acquisition.

[0019] The control and computing center is used to control the mobile platform, sensing module, and communication module.

[0020] The communication module includes 4G / 5G and wireless LAN modules, which are used to transmit inspection data (images, videos, defect reports), device status and alarm information back to the background monitoring center in real time, and receive scheduling instructions, map updates or task replanning instructions from the background.

[0021] The mobile platform uses a wheeled or tracked chassis, possessing excellent track adhesion and obstacle-crossing capabilities. The drive system employs independent hub motors, enabling precise speed and position control.

[0022] The perception module includes positioning and mapping sensors, obstacle detection sensors, and orbital status sensors. The positioning and mapping sensors include a combined navigation system (GNSS / INS) for large-scale global positioning; and lidar (LiDAR) and / or visual odometry (VO) for high-precision continuous positioning and point cloud map construction in areas with poor satellite signal (such as tunnels and canyons).

[0023] Obstacle detection sensors include forward and lateral millimeter-wave radar, ultrasonic sensor arrays, and binocular stereo vision cameras. Millimeter-wave radar is used for mid-to-long-range, all-weather obstacle detection and relative velocity measurement; ultrasonic sensors are used for precise ranging at close range (0.1-3 meters); and binocular stereo vision cameras are used for obstacle classification and recognition (distinguishing between pedestrians, vehicles, falling rocks, etc.) and track feature extraction.

[0024] The track condition sensor includes a high-resolution linear array camera and a laser displacement sensor, used to detect track geometry parameters (gauge, level, elevation) and surface defects (cracks, wear).

[0025] The control and computing center includes a global path planner, a local real-time path planner, and an emergency obstacle avoidance system; The global path planner, based on a pre-stored or real-time constructed track network topology map (including information on tracks, switches, stations, speed-limited zones, and restricted areas), and combined with the set inspection task objectives (such as full-line inspection or inspection of a specified section), uses a global path planning algorithm to calculate a global reference path from the starting point to the ending point. This path is an ordered sequence of path points.

[0026] The global path planning algorithm is the A* algorithm based on a heuristic function. Considering the characteristics of the track network, the heuristic function is designed as a multi-weight cost function, expressed as follows: h(n) = w1 * d(n, goal) + w2 * c_switch(n) + w3 * c_speed(n) Where: d(n, goal) is the geometric distance from the current node n to the target node. c_switch(n) is the cost of passing through a switch. To encourage the path to choose the main route and reduce switch operations, a certain cost is added for each switch the planned path passes through. c_speed(n) is the cost of speed-limited zones. When the path passes through speed-limited zones (such as curves or stations), the time cost is calculated based on the speed limit value and added to the heuristic function to guide the algorithm to prioritize paths that allow high-speed passage.

[0027] Time-based task point planning: When an inspection task involves multiple specific checkpoints that must be passed (such as specific bridges or tunnel entrances), the global planner is upgraded to an optimization algorithm based on the traveling salesman problem. First, the optimal order of visiting the checkpoints is determined. Then, a heuristic function is used to generate sub-paths between every two adjacent points, and finally, these are pieced together to form a complete global task path.

[0028] The local real-time path planner generates a local high-precision occupancy grid map or semantic map centered on the inspection device, which includes static obstacles (such as stopped vehicles and fixed equipment) and dynamic obstacles (such as moving workers and animals), based on real-time environmental perception data, and uses the dynamic window method for real-time trajectory planning.

[0029] The method for real-time trajectory planning using the dynamic window method is as follows: (1) Generation of dynamic window: Establish a dynamic model (v, ω) of the moving platform, where v is the linear velocity and ω is the angular velocity. In the velocity space (v, ω), a reachable dynamic window V_d is determined based on the current velocity, maximum acceleration / deceleration, and control cycle. Only the velocity pairs within the window are sampled.

[0030] (2) Refinement of the multi-objective evaluation function: For each sampled velocity pair (v, ω), simulate its trajectory traj within a short future time interval Δt. Evaluate it using a comprehensive evaluation value G(v, ω): G(v, ω) =α* Heading(v, ω) +β* Dist(v, ω) +γ* Velocity(v, ω) +δ*Semantic(v, ω) Among them: Heading: evaluates the proximity of the trajectory endpoint direction to the local target point (from the global path). Heading(v, ω) = 180° - |θ_traj - θ_goal|, the smaller the difference, the higher the score.

[0031] Obstacle Dist: Evaluates the minimum distance to the nearest obstacle along the entire trajectory. Using a local occupancy grid map, the distance from each pose on the trajectory to the nearest obstacle is calculated, and the minimum value is taken as Dist(v, ω). If the distance between any point on the trajectory and an obstacle is less than a safety threshold, the trajectory is discarded (Dist=0).

[0032] Velocity: Encourages efficient movement. Velocity(v, ω) = v, which is the current sampling linear velocity.

[0033] Semantic Penalty: This evaluation term incorporates the semantic recognition results of obstacles from the visual sensor. Different categories of obstacles generate different penalty weights.

[0034] For dynamic obstacles (such as walking workers), their future positions are either linearly extrapolated or predicted using a Kalman filter. When evaluating a trajectory, not only the current obstacle position but also its predicted position is examined, thus avoiding dangerous behavior of "cutting into" the path of a dynamic obstacle.

[0035] For small, static but surmountable obstacles (such as small stones or tool kits), after visually identifying their category and size, the penalty weight can be appropriately reduced, allowing the device to approach or perform specific obstacle-crossing actions (such as lifting the wheels) at extremely low speeds when necessary, rather than forcing it to detour.

[0036] For critical static obstacles (such as large falling rocks or damaged tracks), a very high penalty weight is assigned to ensure that the planned path must be kept away from them.

[0037] The Semantic(v, ω) term is calculated by combining the degree of intersection between the trajectory and the predicted regions of different types of obstacles, as well as their corresponding weights.

[0038] Trajectory optimization and execution: The velocity pair (v_best, ω_best) that maximizes the evaluation value G(v, ω) is selected from the dynamic window V_d and used as the output of the current control cycle. The underlying controller converts this velocity pair into speed commands for the left and right hub motors, driving the platform movement. This process is repeated in each control cycle (typically 50-100ms) to achieve continuous local path replanning and obstacle avoidance.

[0039] Environmental perception data consists of multi-source data from LiDAR, millimeter-wave radar, and vision cameras, which are fused using Kalman filtering or extended Kalman filtering algorithms.

[0040] When sensors detect a sudden obstacle at extremely close range ahead, and the planning algorithm cannot react in time, the emergency obstacle avoidance system triggers an emergency strategy based on reactive behavior. For example, if an ultrasonic sensor detects an obstacle that appears instantaneously in front of the vehicle, it directly sends an emergency braking command to the underlying controller, which has the highest priority.

[0041] The implementation process of this embodiment is as follows: The integrated navigation system, combined with a pre-installed digital map of the tunnel route, generates an initial path through the tunnel using a global planner. When the integrated navigation system signal fails, the lidar and visual odometry are activated for synchronous positioning and map building. The SLAM process corrects the device pose in real time and constructs a 3D point cloud map including features such as tunnel walls, tracks, and cable trenches. The global path is correlated and fine-tuned in real time on this map. During the inspection process, the forward-facing millimeter-wave radar detects stationary obstacles ahead of the path in real time, and the binocular camera confirms that the obstacles are unstructured. The local planner receives the fused obstacle information and marks it as an impassable area on the local grid map. The local planner uses a dynamic window method, sampling in velocity space, discarding trajectories pointing towards obstacles, and prioritizing trajectories that can safely bypass obstacles and deviate from the global path the least. Finally, a smooth detour trajectory is generated, which is executed by the mobile platform. After detouring, the device automatically returns to the global reference path and continues the inspection. Meanwhile, its track status sensors continuously collect track images and geometric data, and any abnormalities detected will trigger an alarm through the communication module and be uploaded to the monitoring center.

[0042] Environmental perception fusion: The control center receives multi-source data from lidar, millimeter-wave radar, and vision cameras, and performs sensor fusion through Kalman filtering or extended Kalman filtering algorithms to generate a local high-precision occupancy grid map or semantic map centered on the inspection device, which includes static obstacles (such as stopped vehicles and fixed equipment) and dynamic obstacles (such as moving workers and animals).

[0043] Local trajectory planning: Guided by the global reference path, real-time trajectory planning is performed on a local map using the time elastic band algorithm or the dynamic window method.

[0044] The TEB algorithm discretizes the global path into a series of pose points with timestamps, forming a stretchable and compressible "elastic band." Through optimization, while satisfying dynamic constraints (maximum velocity, acceleration) and avoiding obstacles in the local map, the algorithm adjusts the position and time of these pose points to generate a smooth, safe, and executable local trajectory.

[0045] The DWA algorithm samples multiple feasible velocity pairs (linear velocity and angular velocity) in the velocity space, simulates the trajectories of these velocities over a short period, and scores each trajectory using an evaluation function. The evaluation function comprehensively considers: 1) the proximity of the trajectory to the global path; 2) the distance between the trajectory's endpoint and the nearest obstacle (safety); and 3) the trajectory's velocity (efficiency). Finally, the velocity pair with the highest evaluation function score is selected as the current control command.

[0046] Emergency obstacle avoidance strategy: When sensors detect a sudden obstacle at very close range ahead, and the planning algorithm cannot react in time, an emergency strategy based on reactive behavior is triggered. For example, if an ultrasonic sensor detects an obstacle that appears instantaneously in front of the vehicle, it will directly send an emergency braking command to the underlying controller, which has the highest priority.

[0047] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. A track inspection device with autonomous path planning, characterized in that: It includes a mobile platform, a perception module, a control and computing center, and a communication module; the perception module includes a lidar and / or visual odometry for localization and mapping, and millimeter-wave radar, ultrasonic sensors, and visual cameras for obstacle detection; the control and computing center includes a global path planner for generating a global reference path, a local path planner for real-time trajectory planning based on a local environment map fused from multiple sensors, and an emergency obstacle avoidance device for dynamic obstacle avoidance.

2. The track inspection device for autonomous path planning according to claim 1, characterized in that: The local path planning layer uses a dynamic window method to generate a smooth local trajectory while satisfying the dynamic constraints of the mobile platform and avoiding obstacles.

3. The track inspection device for autonomous path planning according to claim 1, characterized in that: When the ultrasonic sensor detects a sudden, nearby obstacle, the emergency obstacle avoidance device sends a braking command with the highest priority to the mobile platform.

4. The track inspection device for autonomous path planning according to claim 1, characterized in that: The global path planner uses a search algorithm to perform global path planning based on the task start point, end point, and track network map, generating a global reference path. The local path planner uses multi-sensor fusion to perceive the real-time environment, constructs and updates a locally occupied grid map, and uses the global reference path as a guide to generate a collision-free local trajectory on the local map based on the dynamic constraints of the mobile platform.

5. The track inspection device for autonomous path planning according to claim 1, characterized in that: The local path planner employs a dynamic window method, and its evaluation function G(v, ω) includes: Heading(v, ω), Dist(v, ω), Velocity(v, ω), and Semantic(v, ω); wherein the Semantic(v, ω) is calculated based on the classification results of obstacles by the visual sensor and the motion prediction of dynamic obstacles.

6. The track inspection device for autonomous path planning according to claim 4, characterized in that: The emergency obstacle avoidance device uses linear extrapolation or Kalman filtering to predict the motion of dynamic obstacles.

7. The track inspection device for autonomous path planning according to claim 5, characterized in that: The classification results of the obstacles generate different penalty weights. For dynamic obstacles, their future positions are linearly extrapolated or predicted by Kalman filtering. When evaluating the trajectory, the current position of the obstacle is checked, as well as its predicted position. For small, static obstacles that can be overcome, their penalty weights are reduced after visually identifying their category and size, allowing the device to approach or perform specific obstacle-crossing actions at extremely low speeds. For critical static obstacles, extremely high penalty weights are assigned to ensure that the planned path must stay away from them.

Images