Research On Collision Detection And Path Planning Methods For Unmanned Vehicles In Dynamic Environments

Driverless cars are gaining more and more attention as one of the ways to improve traffic efficiency and reduce human traffic accidents,and as one of the ways to free human hands and promote the development of artificial intelligence.At present,when using 3D LiDAR for environment sensing in the unmanned field,the problems of large amount of point cloud data,easy to be obscured,and low reflectivity of objects such as glass lead to the lack of real-time and accuracy of 3D LiDAR detection of obstacles.Collision detection is a crucial part of path planning,and the current collision detection algorithm in the unmanned vehicle field is fast in rough detection but not high in accuracy;precise detection requires a lot of calculations and is difficult to meet real-time requirements.In addition,how to smoothly and safely avoid obstacles in unmanned vehicles has been a great concern in the field of unmanned vehicles.Based on the above problems this paper focuses on the three aspects of 3D obstacle detection,collision detection and path planning for unmanned vehicles,and the main work is as follows:(1)Firstly,to address the problem of large blind area generated by single LIDAR due to the large size of unmanned vehicles,this paper designs a dual LIDAR installation layout scheme adapted to this vehicle model,which increases the point cloud density of the key detection area while expanding the LIDAR perception field of view.The road edge detection algorithm based on sampling and segmented multi-spatial feature fusion is proposed for the dual LiDAR fusion point cloud.and the comparison experiments are conducted with other road edge detection algorithms in multiple road conditions such as structured and unstructured straights and curves to verify the efficiency of this algorithm for road edge detection.Secondly,to address the problem of bush false detection caused by the inability to distinguish passable and non-passable areas in the traditional ray-based fast ground point cloud segmentation method,the slope-raybased roadway point cloud segmentation algorithm is improved,and a line fitting ground point cloud segmentation algorithm that fuses road edge points and fast linear search is proposed,which adopts a new and fast way to fuse road edge points;uses dynamic threshold refinement vertical segmentation area;for the lateral search along the angular direction,the traversal starting point is directly determined by the point cloud radius for fast linear search.The superiority of this algorithm in terms of accuracy and efficiency,especially the substantial improvement in detection speed,is verified by comparing with other algorithms in transverse slope,longitudinal slope and curved road conditions.To address the problem of oversegmentation of point clouds caused by the penetrating effect of LIDAR irradiated glass,this paper adds a secondary clustering along the Z-axis based on the radial bounded nearest neighbor algorithm to re-cluster the over-segmented point cloud clusters,and compares the results with the original algorithm to prove that this algorithm can reduce the effect of glass penetration on point cloud clustering and increase the accuracy of point cloud clustering.(2)To address the problem that it is difficult to evaluate the correct position of the obstacle caused by the large area of the point cloud is obscured,this paper uses the corner point-based L-shaped fitting algorithm to construct the obstacle point cloud directional bracket box,compared with the traditional principal component analysis method to construct the effect of directional bracket box.this algorithm fits the bracket box more closely to the boundary of the obstacle,and the tightness is higher.Then,to address the problems of slow speed and redundant projections in the SAT(Separating Axis Theorem)algorithm for intersection testing between oriented wraparound boxes,this paper proposes a fast SAT-based collision detection method for determining the separation projection axis,which quickly determines the potential separation axis by the projection of the geometric center line of two oriented wraparound boxes on each symmetry axis,and then determines whether the two oriented wraparound boxes are intersecting by the separation This method can effectively reduce the number of projections of point lines and improve the detection efficiency.Finally,the detection speed of this algorithm and other collision detection algorithms are compared in two experimental scenarios:the obstacles are randomly distributed on the path of the unmanned vehicle.and the obstacles and the unmanned vehicle are set at multiple relative angles according to a certain rule,which proves that this algorithm can significantly improve the detection efficiency.(3)For the problem that some global path points are lost due to high vehicle speed,this paper adopts segmented three-time Hermite interpolation method to interpolate the path points with excessive distance to complete.Secondly,the local path planning algorithm based on path point sampling is used to generate multiple candidate local paths by dividing the pre-selected global path points into three segments using different sampling methods,and then the local paths are smoothed by gradient descent to obtain curvature-smoothed paths,and the cost function of each path is calculated.In order to reduce unnecessary collision detection and determine the optimal path as soon as possible,this paper separates the collision detection from the cost function and selects the candidate paths for collision detection in the order of low to high cost function score.The Pure Pursuit algorithm is used to control the unmanned vehicle to follow the planned local path.In this paper,an unmanned vehicle simulation environment based on ROS and Gazebo and a simulated unmanned vehicle with Ackermann model are built,and the local path planning algorithm used in this paper is verified in simulation environments such as straight,curved,single obstacle and multiple obstacles.Finally,this paper designs and builds an unmanned vehicle entity,selects,lays out and installs the required sensors,and verifies the perception and planning algorithms proposed in this paper on a real vehicle.The experiment proves that the unmanned vehicle can accurately and quickly perform road edge detection,ground filtering,and 3D point cloud clustering detection when tracking the global path,and according to the obtained obstacle information,the unmanned vehicle can plan a smooth obstacle avoidance path,which realizes the smooth and safe obstacle avoidance of the unmanned vehicle at low speed.