This paper describes the authors' 3-D visualization system that represents traffic activities at an intersection and renders views from arbitrary different viewpoints. Obviously, such visualization system is useful for drivers because they can be aware of what is going on around them. Also, it can be used for post facto inspection from arbitrary viewpoints after traffic accidents. The system can be subdivided into three parts: (1) vehicle detection and tracking, (2) 3-D map modeling and geometric transportation, and (3) view reconstruction from arbitrary view points. In this paper, the authors first describe vehicle detection and occlusion robust tracking algorithm utilizing Spatio-Temporal Markov Random Field model. Secondly, they present how background video-textures and vehicle locations are transformed from input images to 3-D virtual space. Finally, they show some results they have obtained by their visualization system and discuss them.
Abstract