Research On Navigation Methods Based On Multi-Source Perception In Unstructured Environments

Autonomous navigation of unmanned ground vehicle is of great significance to complete complex tasks such as field search and rescue,resource exploration,and regional reconnaissance in unstructured environments.Compared with structured environments such as highway,autonomous navigation faces more challenges in unstructured environments.Currently,terrain traversability mapping method based on single-source perception is difficult to adapt different complex terrains,resulting in a decline in map accuracy.In addition,path planning method based on the binary space division of free space and obstacles lacks consideration of terrain traversability information,which increases the overall traffic difficulty of path and reduces safety of navigation system.In order to solve the above problems,this thesis fuses geometric and semantic information to build traversability maps,and considers terrain traversability score for path planning,aiming to improve the overall performance of navigation system in unstructured environments.The main research content of this thesis is as follows:(1)This thesis integrates camera,Li DAR,GNSS receiver,computer and wirecontrolled chassis vehicle to build a tracked navigation platform to conduct navigation method research,briefly describing vehicle hardware system and software module functions such as tracked vehicle positioning,terrain traversability mapping,global path planning and vehicle motion control.(2)Aiming at the problem that the terrain traversability mapping method based on geometric terrain features lacks terrain semantic information,which leads to decline in map accuracy,a terrain traversability mapping method based on geometric-semantic fusion is proposed,which reduces misjudgment in areas with inconspicuous structural features.This method constructs an environmental height map based on Li DAR data,estimates geometric terrain features according to terrain height,and calculates geometric traversability score by combining the platform traversability analysis result;performs terrain semantic segmentation based on image data,and maps the image semantic segmentation results to traversability map to calculate the semantic traversability score;finally,according to the Bayesian theorem,the geometric and semantic information is fused to judge the terrain traversability.Compared with the method based on geometric terrain features,the method proposed in this thesis improves the average accuracy by 3.9%and the accuracy by 5.74% in actual construction site scene.(3)Aiming at the problem that the path planning method based on obstacles and free space binary map model lacks consideration of terrain traversability information,resulting in overall traffic difficulty,a global path planning method considering traversability score for unstructured environments is proposed.Based on the constructed traversability map,this method maps the precomputed traversability score to the cost to improve the Hybrid A* algorithm,uses fast marching method as heuristic function,designs the cost function to optimize traversability score,and considers traversability score to smooth the paths.The generated path can reach a balance between length of the path and overall traffic difficulty.At the cost of increasing the length of the path,the overall traffic difficulty of the path is greatly reduced.The effectiveness and practicability of the proposed method are verified in simulations based on random terrain map and experiments in actual construction site scene.