Multi-Person Tracking in Smart Surveillance System for Crowd Counting and Normal/Abnormal Events Detection

Automated video surveillance addresses people's real-time observation to describe their behaviors and interactions. This paper presents a novel multi-person tracking system for crowd counting and normal/ abnormal events detection at indoor/outdoor surveillance environments. The proposed system consists of four modules: people detection, head-torso template extraction, tracking and crowd cluster analysis. Firstly, the system extracts human silhouettes using inverse transform as well as median filter reducing the cost of computing and handling various complex monitoring situations. Secondly, people are detected by their head torso due to less varied and hardly occluded. Thirdly, each person is tracked through consecutive frames using the Kalman filter techniques with Jaccard similarity and normalized cross-correlation. Finally, the template marking is used for crowd counting having cues localization and clustered via Gaussian mapping for normal/abnormal events detection. The experimental results on two challenging datasets of video surveillance such as PETS2009 and UMN crowd analysis datasets demonstrate that the proposed system provides 88.7% and 95.5% in terms of counting accuracy and detection rate.