Person specific activity recognition using fuzzy learning and Discriminant Analysis

One of the major issues that activity recognition methods should be able to face is the style variations observed in the execution of activities performed by different humans. In order to address this issue we propose a person-specific activity recognition framework in which human identification proceeds activity recognition. After recognizing the ID of the human depicted in a video stream, a person-specific activity classifier is responsible to recognize the activity performed by the human. Exploiting the enriched human body information captured by a multi-camera setup, view-invariant person and activity representations are obtained. The classification procedure involves Fuzzy Vector Quantization and Linear Discriminant Analysis. The proposed method is applied on drinking and eating activity recognition as well as on other activity recognition tasks. Experiments show that the person-specific approach outperforms the person-independent one.