Visual Localization Based Dense Mapping

Visual Navigation is a hotspot of robotics, among which visual localization, perception and mapping are all the core issues. Vision based localization and terrain perception method possesses the advantages of high accuracy and low power, especially the visual odometry algorithm owns little accumulative error compared with IMU or wheel odometry localization. Hence, this paper proposed a visual localization based dense mapping method and a visual navigation solution. The visual localization adopts binocular motion estimation to obtain global localization and trajectory, which uses SIFT to track features between frames, uses RANSAC to remove mismatching and least square method to obtain motion estimation. This paper also proposed Joint Reliability based Disparity Expansion algorithm to improve existing algorithm's performance. This paper also designs a self-calibration based wide-baseline stereo matching method for long range terrain reconstruction. Moreover, a grid based mapping method is developed to build DEM map and perform the elevation thresholding and optional slope estimation for traversability analysis and traversability map building. Finally, dynamic path planning method is used to guide robot avoid all obstacles and arrive the destination. With series of experiments, the visual navigation solution proposed in this paper is proved to be feasible and robust.