Performance comparison of clustering techniques in detection of dominant flow field in crowded scene

Video scene of real world crowd movement present great challenges for real-time visual surveillance system. This paper presents a framework that detects and cluster dominant crowd flow field into region of interest from a video sequence. A particle dynamic system is applied on the video scene represented by optical flow features obtained based on classical-NL technique. Time integration of the system gives the individual particle motion trajectory. The collection of motion trajectories gives the region of interest. We obtained the segmented dominant flow regions by clustering the region of interest. The clustering of flow field may indicate the possible presence of instability in the segmented regions. Performances of different clustering techniques are compared in this study. Comparison of classification accuracy, precision, recall and F-measures are reported. Different boundaries of flow segments are detected. A change observed on segmented flow field is considered as anomaly. The experiments are carried out on publicly available datasets of crowd.