Memory-based recognition of human behavior based on sensory data of high dimensionality

This paper explores memory-based approaches to the recognition of human behavior that relies on a database of previously categorized instances of sensory data. To overcome the curse of dimensionality, we examine two related methods that both rely on a hierarchical division of the sensory space using a decision tree. The first approach iteratively applies linear discriminant analysis to divide the sensory space in half in order to construct a binary tree for recognizing behaviors. We have verified the effectiveness of this approach for real-time behavior recognition using infrared sensors distributed in a desk environment and compared its results to those of Quinlan´s C4.5. The second approach applies the well-known ID3 algorithm to the construction of a decision tree based on an information criterion. We use it to recognize browsing behavior at a video rental shop. Inferences are derived directly from the binarized pixel data of four wide-view cameras. Both systems offer behavior recognition rates in excess of 90%.