Traversability Analysis Of Quadruped Robot Based On Sparse Point Cloud In Rough Terrain

Compared to the rapid development of self-driving cars in the road environment,the autonomous navigation capability of robots in the field environment still lags behind,and many fields including forestry,agriculture,military,and border patrol are in urgent need of robots that can operate autonomously in complex field terrain.Among mobile robots,quadruped robots have strong terrain adaptability and are an excellent vehicle for field movement.Accurate and complete terrain modeling is a necessary prerequisite for autonomous navigation in the field environment,and the first task is to assess the traversability of the terrain.3D LiDAR is often used to sense the field environment,but the inherent sparsity of LiDAR can lead to incomplete perception.Therefore,it is of great theoretical and practical significance to study the traversability analysis method of quadruped robots for unstructured field environments and to construct accurate and complete traversability maps based on LiDAR in real time.In this thesis,the following research contents are carried out for the two key problems of terrain-dense map building and traversability analysis:(1)For the demand of autonomous navigation in unstructured field environment,build a traversability analysis experimental platform with quadruped robot as the mobile platform and 3D LiDAR as the sensing system.The terrain model based on 2.5D elevation map was constructed by pre-processing the raw LiDAR point cloud with coordinate transformation and downsampling.To address the problem of limited ability of 2.5D elevation map to detect overhanging structures,an overhanging structure judgment mechanism is introduced in downsampling to effectively remove the point cloud corresponding to passable overhanging structures and improve the reliability of the terrain model.(2)In the terrain modeling process,for the problem of sparse and irregular LiDAR point cloud data,a U-Net based Elevation Completion Network(ECN)is proposed to model the elevation completion as an image completion problem by aggregating multiple frames of point cloud data and graying them out.ECN achieves prediction of unknown elevation by learning contextual elevation information.Compared with the Bayesian Generalized Kernel(BGK)inference algorithm,the elevation error is reduced by 27.3%and the running time is reduced by 84.4%,which significantly improves the accuracy and real-time performance of terrain modeling.(3)To address the problem that a single feature cannot accurately describe the traversability of complex field environments,a traversability analysis model with the fusion of multiple terrain features is proposed.The model models the vegetation density based on the penetration rate of LiDAR and optimizes the calculation methods of terrain features such as step height,surface slope and surface roughness,which can achieve accurate description of traversability in the field environment,classification accuracy Acc(Accuracy)reached 72.3%.To meet the real-time requirements,the traversability of each terrain feature is estimated sequentially according to the computational complexity and importance of terrain features,which greatly improves the efficiency of traversability classification.(4)To address the problem of differences in the traversability criteria of different robot platforms,a correlation model between terrain features and robot attributes is proposed to achieve generalization to different robot platforms by using robot platform attributes such as maximum leg lift height and maximum climbing angle as model inputs.Finally,the effectiveness of the traversability building method proposed in this thesis is evaluated in a field environment,and it is verified that the method can be used online in a real environment.