Dominant flow extraction and analysis in traffic surveillance videos

Flow analysis of crowd and traffic videos is an important video surveillance task. In this work, we propose an algorithm for long-term flow segmentation and dominant flow extraction in traffic videos. Each flow segment is a temporal sequence of image segments indicating the motion of a vehicle in the camera view. This flow segmentation is done in the framework of Conditional Random Fields using motion and color features. We also propose a distance measure between any two flow segments based on Dynamic Time Warping and use this distance for clustering the flow segments into dominant flows. We then model each dominant flow by generating a representative flow segment, which is the mean of all the time-warped flow segments belonging to its cluster. Using these dominant flow models, we perform path prediction for the vehicles entering the view and detect anomalous motions. Experimental evaluation on a diverse set of challenging traffic videos demonstrates the effectiveness of the proposed method.