Trajectory Based Activity Discovery

Abstract

This paper proposes a framework to discover activities inan unsupervised manner, and add semantics with minimalsupervision. The framework uses basic trajectory informationas input and goes up to video interpretation. The workreduces the gap between low-level information and semanticinterpretation, building an intermediate layer composedof Primitive Events. The proposed representation for primitiveevents aims at capturing small meaningful motions overthe scene with the advantage of being learnt in an unsupervisedmanner. We propose the discovery of an activity usingthese Primitive Events as the main descriptors. The activitydiscovery is done using only real tracking data. Semanticsare added to the discovered activities and the recognition ofactivities (e.g., “Cooking”, “Eating”) can be automaticallydone with new datasets. Finally we validate the descriptorsby discovering and recognizing activities in a home careapplication dataset.