Section: Application Domains

Human action recognition

We are particularly interested in the analysis and recognition of human actions and gestures. The vast majority of research groups concentrate on isolated action recognition. We address continuous recognition. The problem is difficult because one has to simultaneously address the problems of recognition and segmentation. For this reason, we adopt a per-frame representation and we develop methods that rely on dynamic programming and on hidden Markov models. We investigate two type of methods: one-pass methods and two-pass methods. One-pass methods enforce both within-action and between-action constraints within sequence-to-sequence alignment algorithms such as dynamic time warping or the Viterbi algorithm. Two-pass methods combine a per-action representation with a discriminative classifier and with a dynamic programming post-processing stage that find the best sequence of actions. These algorithms were well studied in the context of large-vocabulary continuous speech recognition systems. We investigate the modeling of various per-frame representations for action and gesture analysis and we devise one-pass and two-pass algorithms for recognition.