Tracking of multiple humans using subtraction stereo and particle filter

This paper proposes a method for the automatic tracking of multiple humans in various scenes using a stereo camera. The proposed method detects candidate regions of humans using “subtraction stereo,” which restricts stereo matching to foreground regions extracted by subtraction and obtains distance information for those regions. Tracking of humans is carried out for the extracted regions using a particle filter. The particle filter consists of four steps, prediction, calculation of likelihood, data association, and resampling. Three features: distance, color, and direction of motion are used in the proposed method to achieve robust human tracking. When humans with similar clothing colors or human direction pass each other, occlusion occurs, and tracking often fails. The proposed method adds robustness to occlusions by explicitly considering the distance, color, and direction of human motion. The proposed method has been evaluated through experiments using a stereo camera that simulates a surveillance camera in real scenes. Tracking accuracy of more than 90% has been achieved in three different scenes, which shows the effectiveness of the proposed method for tracking humans.