An integral stochastic approach to image sequence segmentation and classification

Finding and identifying characteristic or meaningful image sequences in a continuous video stream is a challenging task with many applications. This paper presents a new and efficient approach to these temporal segmentation and classification problems based on hidden Markov models (HMMs). The basic principle consists in continuously observing the output scores of the HMMs at every time step. Peaks, which appear in the individual HMM output scores, allow to determine in an integral way which image sequence occurred at what time. The application of our method to the spotting of connected dynamic hand gestures provided excellent recognition results and a high temporal accuracy.