Action exemplar based real-time action detection

We propose a real-time action detection system based on a novel action representation and an effective learning method with a small training set. We represent actions with a new feature that measures the ¿global¿ distance from a set of action exemplars, where action exemplars are constructed from a vocabulary that encodes ¿local¿ instantaneous body motions. A cascade of linear SVM is used to learn target actions, where at each layer a selective set of exemplars is chosen and variations between locally similar actions are trained in a coarse to fine manner. The method is further extended to incrementally learn a new action with a single example. The method is implemented as a real-time system that can detect actions at frame rate. The performance is extensively validated by evaluating on public and in-house action datasets.