Research On Pedestrian Detection In Static Image

Pedestrian detection is a hot and knotty issue in computer vision. It has widespread application value, such as intelligent video surveillance, advanced driving assistance systems, intelligent robot, etc. Pedestrian detection system in static image mainly includes image preprocessing, feature extraction, classification and non-maximum suppression (NMS). Most studies of pedestrian detection have concentrated on feature extraction, feature learning, classifier and speed up, etc. However, rare attention has been focused on NMS. NMS is an indispensable part of the post-processing. Today, the most common approach of NMS is based on greedy strategy that only uses the information of overlapping area when suppressing other windows. Aiming at this problem, three improvements of NMS were proposed based on ACF (Aggregate Channel Features) detector; Aiming at the problem that most false positives is hard to be suppressed by NMS, a two-stage pedestrian detection scheme is further researched that using detection first then verification. The main contributions in this thesis is as follows:(1) A NMS with the information of scale rate is proposed. It can suppress those false positives which are generated by arm, leg and other columnar-like part of pedestrian, thus reducing false detecting rate.(2) A NMS of keeping the outlying detection windows is proposed. It can reduce false detection rate and miss rate. The decrease of miss rate is mainly due to the increase of detection rate when the foot or head of pedestrian was occluded.(3) A NMS with the capability of revive is proposed. Taking the suppressed detection windows to be suppressed again can increase the probability of fusion outlying detection windows which can revive the suppressed windows.(4) A two-stage pedestrian detection scheme is proposed. In the first stage, we use ACF to generate object hypothesis which is a fast and high precision detector. In the second stage, we verified those hypothesis with DPM that is a detector with the ability of describing global information in some extend.The experiment result, on the INRIA pedestrian dataset, showed that the improved NMS can significantly improve the detection accuracy with negligible increase of time consumption. The two stage pedestrian detection algorithm can remove most of false positives, thus improving the detection accuracy prominently.