Discovering and Describing Activities by Trajectory Analysis

We propose a novel framework which automatically discovers and semantically describes typical activity patterns in surveillance situations, by analyzing object trajectories. In our framework, an activity pattern contains three elements: a) the source, b) the sink and c) the motion pattern. The sources and sinks in the scene are learned by clustering endpoints (origin and destination) of trajectories. The motion patterns in our framework are only related to temporal dynamics and invariant for spatial translation. They are learned through a 2-level hierarchical model. In the first level, the atomic motion patterns are learned by trajectory segmentation and sub-trajectory clustering. In the second level, the motion patterns are learned by clustering the sequences of atomic motion patterns. Combining the learned sources, sinks and motion patterns, activity patterns are easily distinguished and described in natural language. The effectiveness of our method is demonstrated by experiments on vehicle trajectories extracted in traffic scenes.