Human tracking driven activity recognition in video streams

In this paper we describe a method to visually recognize activities from video streams driven by a robust tracking algorithm. The tracking algorithm uses object flow as a motion model that estimates the displacement and the direction of objects in imagery streams; as observation model, object flow is complemented by a generative prior based on the Gaussian Process Latent Variable Model for increased robustness to visual alterations of the tracked object (human). We then use tracking results as input to drive the recognition of activities by hidden Markov models-based classifiers. We evaluate the performance of the examined approach under real-life activity recognition scenarios in visually complex environments characterized by lighting changes, occlusions and outliers.