Measuring Crowd Collectiveness with Trajectory Smoothing

Analyzing videos of crowd scenes plays an important role in surveillance systems. It is essential to measure crowd collectiveness, which indicates the degree of individuals acting as a union. State-of-the-art approaches mainly relied on the similarity of individual behaviors inside and outside neighborhood to quantify the topological structures of collective manifolds of crowd. Individual behaviors were estimated using velocity of individual trajectories tracked in consecutive frames. However, trajectories could be unstable due to the variations of speed and moving direction of individuals at different tracking duration, hence inaccurate collectiveness measurement. In this paper, we propose an approach to enhance the accuracy by stabilizing motions of individuals. The stabilization process is performed using smoothing methods. We evaluated our methods on a benchmark dataset, Collective Motion. Experimental results indicate that smoothing-based approaches outperform the original state-of-the-art approach.