High performance pedestrian detector using local segmentation self-similarity in complex scenes
Hongbo Gao, Hongyu Wang, Xiaokai Liu, Xiaorui Ma · Pattern Recognition and Image Analysis · 2014
Although a variety of promising approaches exist, it is still a hard work to obtain desirable results in the area of pedestrian detection, especially in crowded and cluttered scene. In this paper, we present a detector which includes a discriminative shape descriptor—Local Segmentation Self-Similarity (LSSS) and induces a simple but sophisticated sample strategy. The descriptor represents the local shape of the object based on saliency on log-polar coordinate. The image is divided into disjoint cells, and the AdaBoost algorithm is adopted to integrate the local shape feature into a simple and powerful classifier. In detecting step, a greedy procedure is utilized for eliminating the repeated detections via non-maximum suppression. Experiments show that our approach achieves the considerable improvements in dealing with heavy occlusion and mutative background.