Research On Visual Navigation Techniques For Intelligent Robots In Unknown Environments

Intelligent robot has already been one of the academic frontiers and focuses both in domestic and overseas with the development of science, technology and industrial automation. Autonomous navigation technique, which contains scene capturing, processing, analyzing, understanding and decision stages, is the one of the most important properties for a highly intelligent robot. Visual sensor has been one of indispensable parts for environmental perception since it has many advantages, such as large scene information, large detection range and abundant image features. The study of visual navigation is of important theoretical and practical significance. Visual navigation has promising future in the industrial manufacturing, national defense and service industry. Nowadays, great achievements have been made in dealing with autonomous navigation tasks in known and structured environments. However, it has not been settled entirely in unknown and unstructured environments. Therefore, in order to satisfy the requirement of autonomous navigation in unknown and cluttered environments, this dissertation focuses on several key issues of visual navigation, such as obstacle avoidance, object detection and tracking, etc.(1).A dense matching algorithm based on the similarity of local image descriptors is researched. Especially, two dense matching methods based on passive and active vision, which are suitable for different occasions, are proposed as follows.Firstly, a novel matching-cost filtering model for dense matching is proposed. It is based on an edge-preserving filter of which the adaptive support weights are computed using a hierarchical clustering algorithm. As a result, most of the noise is eliminated and the depth discontinuity boundaries are preserved fairly well in the disparity maps. Furthermore, the proposed filter reduces the complexity from O(N) to O(N), where N is the number of pixels within the whole image. The experimental results have demonstrated that this method outperforms other local algorithms in terms of both speed and accuracy. The effects of parameter settings on accuracy and efficiency are also discussed. This study offers recommendations for future practical applications.Secondly, an adaptive-window matching method is presented to find an optimal matching window for the laser spot patterns,which are generated by two laser spot projectors. The integral grayscale variance and integral gradient variance are presented as for the quality assessment of image texture. The experimental results for the practical conditions have demonstrated that this approach can be capable of high-speed processing compared to other methods. Moreover, it offers high-quality disparity maps for dense stereo correspondence, especially for the pixels near the self-occlusion and depth discontinuities.(2).An invariant Hough random ferns (IHRF) in image is proposed. It mainly includes rotation and scale invariance feature, codebook generation, random ferns classifier training, Hough map generation, Hough voting and searches for the maxima. The detection results show that the proposed method is robust to rotation and scale variations, changes in illumination, part-occlusions and non-rigid transformations. The results of performance tests provide effective guidelines for parameter selections for IHRF.In addition, a further study applies the IHRF to the problem of RGB-D based object detection. Compared with the traditional3D local feature extraction techniques, this method is an effective way to reduce the computations for feature detection and matching procedures. Moreover, the proposed algorithm which combines2D image and3D depth image has improved the accuracy compared with the object detector in2D image.(3).According to the demand of processing speed and accuracy for practical applications, a multi-object tracker based on bounding-box detector and optical flow, and an online tracker based on part-based detector and segmentation are proposed respectively. The first approach consists of three important components:the compositional multi-feature containing colour and gradient information, the cascaded classifier based on Random ferns classifier and K-Nearest Neighbor algorithm, and the tracking framework based on optical flow and multi-object detector. In the experiments, it is obviously found that the proposed method can be used to address real single-or multi-object tracking tasks under several difficult issues, such as part-occlusions, changes in illumination and cluttered backgrounds, etc.Secondly, an online tracking approach based on an efficient clustering scheme is proposed, which is able to prevent drifting problem. For local codebook generation, it groups the binary masks according to the geometric relations of object parts; for template segmentation process, it classifies image patches by spatial distribution and appearance similarity. This top-down segmentation delivers a more precise description of the object and is used to decrease the noise in the online learning stage for object tracking. For the online object tracking task, this method has been validated using several datasets under challenging conditions, such as cluttered background, changes in posture, and non-rigid deformations. In fact, the proposed method is a stable and effective way to avoid the drifting problem.(4).A visual navigation system for intelligent robot is designed. Several tests for the visual navigation and positioning are presented, such as obstacle avoidance, object detection and object tracking. The experimental results show that all of the proposed methods are effective and feasible.