Stereo-based pedestrian detection using the dynamic ground plane estimation method

Pedestrian detection requires both reliable performance and fast processing. Stereo-based pedestrian detectors meet these requirements due to a hypotheses generation processing. However, noisy depth images increase the difficulty of robustly estimating the road line in various road environments. This problem results in inaccurate candidate bounding boxes and complicates the correct classification of the bounding boxes. In this letter, we propose a dynamic ground plane estimation method to manage this problem. Our approach estimates the ground plane optimally using a posterior probability that combines a prior probability and several uncertain observations due to cluttered road environments. Our approach estimates a ground plane optimally using a posterior probability which combines a prior probability and several uncertain observations due to cluttered road environments. The experimental results demonstrate that the proposed method estimates the ground plane robustly and accurately in noisy depth images and also a stereo-based pedestrian detector using the proposed method outperforms previous state-of-the art detectors with less complexity.