Fast depth estimation using spatio-temporal prediction for stereo-based pedestrian detection

Generating a high-quality disparity map is a fundamental step in many applications such as stereo vision-based pedestrian detection for advanced driver assistance systems (ADAS). One of the major challenges in generating accurate depth maps is the huge computational complexity in stereo matching. In this paper, we propose a fast depth estimation technique for real-time applications. The proposed architecture employs spatial and temporal disparity prediction modules in order to decrease spatio-temporal redundancy. In order to evaluate the performance of the proposed method systematically, we apply the generated depth maps to a stereo-based pedestrian detection system. Simulation results show that the proposed method reduces the computational complexity by 68%-83% while maintaining comparable detection performance with the full-search block matching algorithm used as a reference.