Integrated Tracking and Recognition of Human Activities in Shape Space

Abstract

Activity recognition consists of two fundamental tasks: tracking thefeatures/objects of interest, and recognizing the activities. In thispaper, we show that these two tasks can be integrated within theframework of a dynamical feedback system. In our proposed method, therecognized activity is continuously adapted based on the output of thetracking algorithm, which in turn is driven by the identity of therecognized activity. A non-linear, non-stationary stochastic dynamicalmodel on the ``shape{''} of the objects participating in the activitiesis used to represent their motion, and forms the basis of the trackingalgorithm. The tracked observations are used to recognize the activitiesby comparing against a prior database. Measures designed to evaluate theperformance of the tracking algorithm serve as a feedback signal. Themethod is able to automatically detect changes and switch betweenactivities happening one after another, which is akin to segmenting along sequence into homogeneous parts. The entire process of tracking,recognition, change detection and model switching happens recursively asnew video frames become available. We demonstrate the effectiveness ofthe method on real-life video and analyze its performance based on suchmetrics as detection delay and false alarm.