Analysis of Crowded Scenes using Holistic Properties.pdfVIP

Appears in 11th IEEE Intl. Workshop on Performance Evaluation of Tracking and Surveillance (PETS 2009), Miami, 2009. Analysis of Crowded Scenes using Holistic Properties Antoni B. Chan Mulloy Morrow Nuno Vasconcelos Department of Electrical and Computer Engineering, University of California, San Diego abchan@ucsd.edu, mmorrow@ucsd.edu, nuno@ece.ucsd.edu Abstract modeling the crowd flow in the scene using the dynamic texture model, and training event classifiers on dynamic tex- We present results on the PETS 2009 dataset using tures [2]. In the remainder of this paper, we review the surveillance systems based on holistic properties of the crowd counting system from [1] in Section 2, and the dy- video. In particular, we evaluate a crowd counting sys- namic texture classifiers from [2] in Section 3. We then dis- tem, based on regression of holistic (global) features, on the cuss the experimental setup and results on the PETS 2009 PETS 2009 dataset. We also present experimental results dataset in Section 4 and 5. on crowd event detection when using the dynamic texture model to represent holistic motionfl ow in the video. 2. Crowd counting using low-level features We adopt the crowd counting system proposed in [1], 1. Introduction which is based on regression of low-level features. Con- sider the example scene shown in Figure 1, where the goal is There is c