We propose a framework for detecting, tracking and analyzingnnon-rigid motion based on learned motion patterns. The frameworknfeatures an appearance based approach to represent the spatialninformation and hidden Markov models (HMM) to encode the temporalndynamics of the time varying visual patterns. The low level spatialnfeature extraction is fused with the temporal analysis, providing anunified spatio-temporal approach to common detection, tracking andnclassification problems. This is a promising approach for many classesnof human motion patterns. Visual tracking is achieved by extracting thenmost probable sequence of target locations from a video stream using ancombination of random sampling and the forward procedure from HMMntheory. The method allows us to perform a set of important tasks such asnactivity recognition, gait-analysis and keyframe extraction. Thenefficacy of the method is shown on both natural and synthetic testnsequences
展开▼