Research On Motion Planning And Image Understanding For Autonomous Mobile Robot

Autonomous mobile robot embodies a comprehensive intelligence covering multiple subjects, so related research and applications have attracted increasing attention. Robot vision is a challenging issue in the field of robotics. Research on it is divided into three levels: perception level, application level and cognitive level. Corresponding to the three levels, this dissertation focuses on stereo matching algorithm, motion planning method, and image understanding solution. First of all, this dissertation describes progresses in stereo matching, motion planning and image understanding, and presents key technical issues and development trends. And then, the main contributions are summarized from three aspects. Towards matching and perception, the first is to propose a stereo matching algorithm, which gives consideration to matching efficiency and performance. Based on this, an obstacle detection method is designed. Towards planning and behavior, the second is to propose a hybrid interaction mechanism and the the techniques implemented in each module. Towards reasoning and recognization, the third is to construct a fused feature and then to propose a cascaded framework of inference. Finally, extensive practical experiments are used to verify the proposed methods.From the perspective of perception intelligence, a hierarchical stereo maching algorithm for obstacle detection in the pyramid disparity-offset space is presented. An adaptive dual weighted cost aggregation is designed to improve the rationality of cost calculation. Based on this, bidirectional dynamic programming with multiple transitions structure is presented to integrate the inconsistency in recursive processes. In the forward or backward step, horizontal and vertical optimizations are considered simutanously, and multiple almost-optimal transitions structure is used to multi-candidate backtrack. In the optimization process, ground control points are not only adopted, but both global information of multiple scanline constraints and punitive and incentive measures are integrated into a unified energy function. A series of experiments and comparisons verify both accuracy and efficiency of our algorithm. Furthermore, a two-stage approach with perception-confirmation structure is proposed to detect obstacles. Experiments based on real robot platform in different environments validate effectiveness and practicability of the proposed method.From the perspective of behavior intelligence, an interactive mechanism and modular approaches are proposed for hybrid motion planning in unknown, dynamic and cluttered environments. A bidirectional interaction is designed by deliberative candidates negotiating with the feedback on reactive evaluation. In the deliberative module, a multilayer structure and a switching mode in control sets are designed to construct a concise and flexible state lattices. Furthermore, a multitask-parallel algorithm is proposed to heuristically construct a search tree of the reachable graph to improve search efficiency. In the reactive module, a hierarchical structure is designed to integrate the reaction optimization and situation-dependent adjustment. Based on manifold correlations, piecewise criterions rather than a single function are proposed to cater to different stages of planning. Extensive experiments using Pioneer 3DX platform verify the efficacy, reliability and robustness of our approach in complex environments.From the perspective of cognitive intelligence, a hybrid fused feature detector and cascaded discriminative framework are proposed for image understanding. A local invariant feature is extracted by multiscale and multi-orientation description and dimension reduction. Based on this feature, in serial fusion mode, multiscale histogram of gradient is integrated to comprehensively characterize the local appearance description. Subsequently, in parallel fusion mode, spatial texture structure is integrated to construct a hybrid feature description. Furthermore, in terms of reasoning using discriminative model, a cascaded conditional random field is presented. Some nodes of low layer are adaptively aggregated to form the node of high layer using the trained partition model. This structure can represent local spatial relationships of different levels of components. In addition, the confidence set in low layer is added to the input of node in the high layer. Moreover, the pairwise potential of high layer incorporates contextual information, including coocurrence, relative spatial location, and relative scale, simultaneously. From multi-class object detection and segmentation aspects, extensive experiments in real scene images and comparison with representative methods in PASCAL VOC Challenge verify that our method achieves significant improvement.