Trajectory Based Activity Discovery

This paper proposes a framework to discover activities in an unsupervised manner, and add semantics with minimal supervision. The framework uses basic trajectory information as input and goes up to video interpretation. The work reduces the gap between low-level information and semantic interpretation, building an intermediate layer composed of Primitive Events. The proposed representation for primitive events aims at capturing small meaningful motions over the scene with the advantage of being learnt in an unsupervised manner. We propose the discovery of an activity using these Primitive Events as the main descriptors. The activity discovery is done using only real tracking data. Semantics are added to the discovered activities and the recognition of activities (e.g., "Cooking", "Eating") can be automatically done with new datasets. Finally we validate the descriptors by discovering and recognizing activities in a home care application dataset.