Wide baseline stereo object matching using minimal cost flow algorithm

Monocular vision-based vehicle detection is a low-cost solution for active safety and driver assistance systems (ASDA). However, the depth estimation deviates its true value when the flat ground assumption does not hold. In this paper, we propose a stereo approach with a large baseline to address the issue without extracting three-dimensional features from disparity map. The proposed system first searches vehicle template among possible discrete rectangle boxes in the image pair. The system detects the presence, and estimates the distance of a vehicle simultaneously. This joint problem of detection and matching can be formulated as a minimal cost flow problem, which can be solved efficiently. The experimental results show that not only we have a redundant monocular vision system, but also the performance of both detection and range estimation is significantly enhanced.