Research On Modeling Of Terrain Geometry And Mechanical Properties Based On Planetary Rovers' Visual Information

In the autonomous planetary exploration system,terrain sensing is a crucial link between the external environment and the internal control.It provides fundamental terrain information for multiple tasks such as path planning,virtual simulation and motion control.With the increasing difficulty of Mars exploration missions,more and more applications take terrain mechanical properties into consideration when analyzing impacts on the mission process.Based on the characteristics of Martian terrain and the requirement of terrain-related applications,a new map model with fundamental terrain geometric and mechanical properties is proposed.The map is analyzed and studied in the following three aspects: how to characterize mechanical properties,how to estimate terrain mechanical parameters remotely with vision information and how to mapping with geometric and mechanical properties.This paper presents a generalized map representation that consists of geometric and mechanical properties considering completeness,uniformity and simplicity.Based on the analysis of the terrain bearing and shearing models,the specific parameters of mechanical properties in the map and its corresponding usage in terramechanics-based applications are determined.With the analysis of sensitivity of each parameters,the equivalent stiffness modulus and the equivalent friction modulus are selected ad the dominant parameters which should be estimated dynamically to characterize pressure and friction respectively.A wheel-terrain interaction model is established and the identification of dominant parameters is carried out,which provides data for vision-based dynamic estimation of dominant parameters.To solve the vision-based terrain mechanical properties estimation problem,a one-stage scheme and a two-stage scheme are proposed.In the two stage scheme,terrain segmentation and mechanical properties regression are well studied with concrete implementation.The fully convolutional network and encoder-decoder network are trained and tested on the Mars terrain segmentation dataset.Compared with performance,the fully convolutional network is selected for the mapping experiment.Taking semantic segmentation results as prior knowledge,a probabilistic approach to estimate terrain mechanical properties is adopted and the safety of the estimation is evaluated based on hazard awareness.Learning from the semantic mapping process,a mapping scheme with terrain mechanical properties estimation is designed.The implementation and details of terrain geometric feature layer and terrain mechanical feature layer are elaborated.Based on the proposed mapping scheme,experiments building terrain maps with geometric and mechanical properties are performed in an analogue Martian test yard,which verified the feasibility of the mapping scheme.The validity of the constructed terrain model is verified by high-fidelity simulation.In this paper,the parametric characterization method of terrain geometry and mechanical properties is proposed,and the terrain mechanical parameters are estimated based on vision information.It is successfully applied to the terrain mapping scheme which builds a map with terrain geometric and mechanical properties and provides a terrain perception solution for rovers in the planetary environment.