Depth estimation from a single image in pedestrian candidate generation

Depth estimation, which is mostly performed by stereo vision, is a remarkable task in vision and scene understanding. In this paper, depth map estimation from a single image is investigated and applied in pedestrian candidate generation. To recover accurate depth map from a single image, a Markov Random Field (MRF) model that incorporates both image depth cues and the relationships between different parts of the image is employed. The MRF model can be trained via supervised learning. Then a method is proposed to generate pedestrian candidates using both our estimated depth information and geometric information achieved from the image. Both representations of the scene are fused to limit the region of interest to objects standing vertically on the ground and having certain height. The proposed algorithm is tested using a public database and a considerable reduction in the number of candidate windows is achieved, which translates into a significant time-saving.