Obstacle localization in 3D scenes from stereoscopic sequences

A method for 3D scene segmentation from stereoscopic image sequences registered by a moving camera system is proposed. The method identifies 6DoF ego-motion parameters first. Then, a scene model is built by projecting a triangular mesh onto the dense disparity maps. The triangles are clustered so forming surfaces of separate scene objects that are assigned to obstacle category or context scene objects such as walls depending on their location versus the observer. The method was tested on real stereoscopic scenes and validated on OpenGL generated 3D virtual scenes. Its advantages are: low computing demand and capability of identifying location and shape parameters of the selected obstacles. The developed system is aimed to serve as a vision module part to the travel aid prototype system for the blind.