Automatic semantic labelling of human motion activity
Animesh Garg, A. J. Naftel · 2006
Motion activity analysis has been the focus of much recent research in visual surveillance. However, there has been little work on automated extraction of generic descriptors from video clips and most reported techniques are location or activity specific. This paper proposes a novel approach to semantic description of human activity based on motion trajectory data. Motion is described in natural terms that a human operator would use, such as turning left or turning back. A decision tree classifier is used to learn the mapping between a symbolic trajectory approximation and semantic directional labels. We show that this mapping can be applied across different camera locations and can be used with or without site-specific information to build complex activity descriptions. Preliminary experiments performed with the CAVIAR test scenarios show classification accuracy rates of 89% are achievable.