Dense Stereo-based Real-time ROI Generation for On-road Obstacle Detection

The use of 3D visual information has become widespread as an essential cue for detecting on-road obstacles in ADAS. In this paper, we propose an accurate dense-stereo-based system for the generation of on-road obstacle ROLs. To balance the concerns of computation overhead and algorithm accuracy, this paper presents an efficient depth map generation that combines global stereo matching with depth up-sampling. The entire system has been implemented with a hardware and software partitioning method running on an FPGA and embedded CPU for real-time processing. The implementation results verify that the proposed stereo vision system efficiently outputs accurate ROI candidates for on-road obstacle detection.