high performance computing on graphics processing units: hgpu.org

This paper describes a topological SLAM system using a purely vision-based approach. This robot utilizes a GPU-based omnidirectional catadioptric stereovision system to perceive and plan its path in the environment. Subsequently, the omnidirectional images generated are used to incrementally build a database of image signatures based on the standard 2D Haar Wavelet decomposition. In order to maintain a globally consistent topological map, a relaxation algorithm, which requires local metric information between nodes, is employed each time the appearance-based localization system revisits an existing node in the topological map. The relative transformation of the current position of the robot with respect to the actual position of the matched node is recovered by using a least squares estimation of the transformation parameters of two 3D point patterns generated by the stereovision system. In addition, local metric information is obtained by using the proposed visual odometry system which combines distance measurements calculated by using optical flow techniques which estimates the movement of a web camera relative to the ground being observed and bearing estimates from the omnidirectional catadioptric vision system. Experiments were conducted in a variety of environments ranging from indoor to outdoor environments which demonstrate the feasibility of this approach.