A multistage filtering for detecting group in the crowd
Qian Zhao, Jie Shao, Yan Qing Zhao · 2016
An unsupervised approach of the automatic collective motion group detection is proposed in this paper, which firstly extracted foreground-keypoints based Gaussian Mixture Models (GMMs) and Kanade-Lucas-Toma (KLT) trackers, then foreground-keypoints are multistage filtered by temporal and spatial distance adjacent filter operator, velocity angle filter operator and velocity correlations filter operator between neighboring keypoints to achieve detect group. Our method need no prior information and the number in the K-NN set is self-tuning according to the nearest distance neighboring keypoints in foreground fields. Qualitative and quantitative analysis are carried out on the results of our group detection with other typical methods, experimental results show that our algorithm is more efficient and more robust in complex crowded motion scenes.