中国科学院西安光学精密机械研究所机构知识库

Detecting pedestrians' spatial locations is a key, yet challenging task in video surveillance. From video sequences, we can apply motion detection algorithm to detect the whole pedestrians, and recognize pedestrians' foot-point locations with foot-point recognition method. While we propose a solution to recognize pedestrians' head-point locations for partially occluded pedestrians in complex scenes. Then, a few simple mappings can be used for converting head-point to foot-point and converting 2-D locations to 3-D spatial locations in computer vision. In this paper, we present a pixel-level background sample set motion detection approach based on Self-Balanced SENsitivity SEgmenter, coined SuBSENSE algorithm. Instead of using the same background/foreground segmentation criterion for low and high brightness distribution areas, we use completely different segmentation criterion for low and high brightness, respectively. Besides, for an actual scenario with camouflaged foreground objects, simple color and texture feature could not detect these motion objects. To best address these disadvantages, we introduce normalized color feature and extended local binary similarity pattern (ELBSP) operator by adaptive threshold to segment motion objects for high brightness while providing normalized color feature by perception-inspired confidence interval for high brightness. Due to the diversity of camera gesture in video images, we can't directly gain foot-point and head-point location from the results of motion detection. For foot-point, principal component analysis is employed in getting pedestrian's upright direction and mapping the whole object to this direction. Moreover, color feature, area feature, and position feature are utilized for detecting head-point. Experiments show that it outperforms original motion detection approach and several state-of-the-art methods, and can accurately obtain pedestrians' 2-D locations in real scenarios. © 2017 IEEE.