Spatiotemporal Moving Crowd Tracking via Integral Optical Flow

Abstract

This paper presents a novel approach for tracking moving crowds. Departing from conventional methods that focus on individual pedestrians, our method conceptualizes a moving crowd as a single, dynamically evolving entity. This entity can split into smaller sub-crowds or merge with others to form larger aggregations, making the approach particularly suitable for highly crowded scenarios. The proposed framework operates in two primary stages. First, moving crowds are detected using an integral optical flow technique, which accumulates optical flow vectors across consecutive video frames. Second, crowd identities are maintained via an ID management mechanism underpinned by a contribution matrix. This matrix records the contribution degree of detected crowds in the previous frame to those identified in the subsequent frame. The method is evaluated on manually annotated clips from three publicly available videos. The evaluation yields an average Multiple Object Tracking Accuracy (MOTA) of 0.361. Furthermore, the method demonstrates high performance in capturing crowd dynamics, with average precision and recall for crowd merging reaching 0.942 and 0.811, respectively, and for crowd splitting reaching 0.905 and 0.952, respectively. Additionally, the study defines internal motion patterns, referred to as "groups", within the moving crowds. These groups are identified based on local motion feature similarity and can be tracked in a manner analogous to the crowds themselves. Finally, several parameters are proposed, which hold potential for enabling more in-depth analysis of crowd movement behaviors.