Human Activities at Different Levels of Detail

Classification and recognition of human activity is an important field od computer vision based human motion analysis. The main problem of human motion analysis is the great complexity of human motion, which is rooted in the complicated structure of human body itself. In this paper we explore an alternative view of human motion analysis, which is based on the concept of human motion scale. This concept allows analysis of important issues in computer vision based human motion analysis that have not been sufficiently explored in the related work. In this paper we focus on one particular issue: the appropriate level of detail for a particular human activity classification task. For our experiments we used a video database of recorded squash matches, which have been processed using simple blob tracker under human supervision to obtain positions of players during the whole match. Digitized recordings were annotated by sports experts to provide data about two types of player activity. Annotations are used as a ground truth in our experiments. Results confirm that the choice of the level of detail on which human motion is observed may play crucial role in the efficiency of the classification algorithm. Additionally, we show that our findings are consistent over the whole database of 18 sets of squash play, which were recorded at different tournaments with different players.