Research On Martian Hazard Detection Method

Autonomous Hazard Detection and Avoidance(HDA)has been one of the key tech-nologies for Mars landing missions.The implementation of such technology has a direct and significant impact on safe landing sites selection,as well as subsequent scientific in-vestigation.Therefore,it is an urgent issue to develop autonomous HDA techniques.With the support of Major State Basic Research Development Program of China(973 Program)"Research on navigation,guidance and control for precise planetary landing",the over-all goal of this thesis is to unravel some of the key techniques in HDA for Mars landing missions.For the purpose of developing new theory and methodology in the context of engineering applications,this dissertation brings forward a research on active and pas-sive version-based hazard detection techniques.The main contributions of this thesis are summarized as follows:Terrain surface reconstruction and corresponding hazard detection methods are stud-ied.1)To handle the difficulty of datum surface modeling caused by topographic re-lief,a novel terrain reconstruction method is proposed through multi-scale Thine Plate Spline(TPS)interpolation.The new algorithm is capable of maintaining high reconstruc-tion precision against substantial hazard abundance.With the assumption of local mini-mum as ground(control)point,the proposed method builds a hierarchical structure of such points by down-sampling.Obtain the TPS surface derived from lowest-resolution ground point set and update the control points on each scale.The final terrain datum surface is obtained from the bottom updated ground points.2)Considering the comparability be-tween range error and lander's capability of rock and slope,a novel adaptive threshold estimator is proposed based on the strategy of clustering.Gaussian Mixture Model is ap-plied to build the distribution of residual data,and inner scale estimator is then derived by such data-driven design,enabling automatic regulation.3)Due to the weak capability of safe-danger description of binarized hazard map,a continuous hazard characterization function is designed.Compared with the binarized map,the proposed method can not only judge whether a potential landing area is safe or not,but also how much 'safer' it is.Simulation results demonstrated the effectiveness of the proposed algorithms.A rapid and robust terrain plane fitting technique is studied.For relatively planar or small-sized areas,the key technology of hazard detection is to robustly fit the datum plane in real time under unknown hazard abundances.We propose an adaptive residual scale factor with rapid convergence rate.RANSAC algorithm is utilized to obtain a rough-point set,which is updated with the proposed scale factor to obtain conservative interior-point set.The fitted plane generated by the updated set can significantly remove most of the hazards points,ensuring robustness under varying hazard abundances.A concise and efficient rock detector is studied.Fully considering the limitation of current on-board memory and CPU processor,a region-level image processing strategy is adopted instead of the current pixel-level detection framework.A superpixel-based rock detection approach is proposed through region contrast,converting the rock detec-tion to foreground region enhancement.Only few features are utilized to build region contrast model,in order to meet the constraint of on-line application.Superpixel seg-mentation significantly reduces the memory footprint and computation time.Simulation results demonstrate that our algorithms can generate consistent and desirable performance rapidly.Considering the improvement of spaceborne computing capability in future plan-etary missions,we propose a multi-scale contrast fusion strategy to eliminate the potential rock boundary errors due to limited superpixel number.The simulation shows significant improvement of rock detection precision with the proposed contrast fusion strategy.Multi-feature based rock detection techniques are studied.A multi-dimensional statis-tic feature descriptor is built with fair capability of distinguishing rock and soil.With the unsupervised learning method,background soil model is established by following two strategies:1)PCA-based(Principal Component Analysis)feature selection;2)Sparse-based selection of labeled samples.Compared with the limited-features model,the multi-dimensional redundant model significantly promotes the rock/soil contrast and detection precision.Based on it,a least square operator is carried out to further optimize such con-trast,and Bayesian inference is applied to fuse the optimized detectors.Simulation results demonstrate the validity of proposed algorithms.