Research And Implementation Of UGV Navigation Approach Based On Traversable Area Prediction

The rapid development of robot technology has brought great convenience to people’s production and life,and robot navigation technology plays a crucial role as a core driver in many practical applications.For different application scenarios,robot navigation technology presents diverse development trends.Early navigation technology focused on indoor mobile robot navigation applications,forming a navigation framework using a single-layer LiDAR as the sensor and combined with occupancy grid maps.In recent years,with the increasing demand for autonomous driving,the outdoor highway-level navigation framework based on highdefinition maps and visual perception has become the research focus.Unmanned ground vehicles(UGVs)have broad application potential in areas such as industrial production,logistics distribution,and safety inspection.However,due to the complexity and diversity of application environments,existing technologies still face certain challenges and gaps in meeting their practical application requirements.On the one hand,the outdoor environment is more extensive and complex,and the positioning and perception of unmanned ground vehicles cannot rely solely on single-line laser radar and grid maps,but are more similar to autonomous driving technology.On the other hand,UGVs require more flexible movement routes and cannot drive only along lanes like autonomous cars.However,there is a lack of environment-passable description methods like indoor grid maps in outdoor environments.Therefore,outdoor UGV navigation technology still faces certain challenges.Based on the above issues,this thesis proposes a multi-level planning UGV navigation method based on LiDAR perception.For the perception part.firstly,this thesis proposes an UGV traversability representation called "Circular Accessible Depth",which uniformly represents the traversability of the surrounding environment of the UGV as the maximum accessible distance in each direction centered on the UGV.Secondly,to address the problem of large blind spots of LiDAR,a multi-frame point cloud fusion module based on attention mechanism is designed,which supplements the blind spots while avoiding the interference of residual information of dynamic objects in historical point clouds on prediction results by fusing multi-frame point clouds information.In addition,this thesis also designs a model training process based on semi-supervised learning for the characteristic representation of Circular Accessible Depth,which improves the generalization performance of the perception method.For the planning part,this thesis proposes a multi-level path planning method consisting of global road planning,intermediate-level planning,and local path planning.Among them,global road planning is similar to outdoor road-level planning for autonomous driving,which is used to specify the approximate running direction of the UGV,while local path planning adopts the planning method of indoor mobile robots to control the movement of UGVs.To address the problem of the lack of traversability information for UGVs in outdoor environments,an intermediate-level costmap is established using the traversable information predicted by the perception module,and an intermediate-level path planning is conducted using graph search,making the movement of UGVs more flexible and safe.This thesis is deployed on a real UGV platform and tested in actual environments.Through various comparative experiments and ablation tests,the feasibility of the proposed method in real UGV navigation applications is demonstrated,and it outperforms existing technologies.