Understanding road scenes using visual cues and GPS information

Understanding road scenes is important in computer vision with different applications to improve road safety (e.g., advanced driver assistance systems) and to develop autonomous driving systems (e.g., Google driver-less vehicle). Current vision-based approaches rely on the robust combination of different technologies including color and texture recognition, object detection, scene context understanding. However, the performance of these approaches drops-off in complex acquisition conditions with reduced visibility (e.g., dusk, dawn, night) or adverse weather conditions (e.g., rainy, snowy, foggy). In these adverse situations any prior information about the scene is relevant to constraint the process. Therefore, in this demo we show a novel approach to obtain on-line prior information about the road ahead a moving vehicle to improve road scene understanding algorithms. This combination exploits the robustness of digital databases and the adaptation of algorithms based on visual information acquired in real time. Experimental results in challenging road scenarios show the applicability of the algorithm to improve vision-based road scene understanding algorithms. Furthermore, the algorithm can also be applied to correct imprecise road information in the database.