Space-Time Analysis of Dynamic Scenes

Lihi Zelnik-Manor

The main objective of the presented research is to analyze video data of dynamic scenes based on their behavioral content. We first extract independent space-time components, and then analyze and classify them. Each spatio-temporal component spatially corresponds to a scene component having consistent motion (e.g., a static background scene versus a foreground moving object), and temporally corresponds to consistent behavior over time, i.e., a single action or event. For example, if a person first walks and then starts running, this corresponds to two different temporal events, hence two different temporal components. The proposed methods use only temporal information cues and ignore appearance effects.

The extraction and classification of independent space-time components is done using two different approaches. The first approach, which we refer to as the "factorization approach" or the "subspace-based approach", is geometric in its nature and is based on linear subspace constraints. The second approach, to which we refer as the "statistical approach", is probabilistic and regards an event as a stochastic process. These approaches and their applications will be presented in this talk.

Maintained by Philippos Mordohai