Abstract
There is growing interest in automatic recognition of human actions in video sequences shot by surveillance cameras. However, it's difficult to analyze human actions in real environments. That is, almost all of the current techniques can only detect simple actions within video sequences showing controlled environments. We propose action recognition methods based on multiple trajectories that can identify human actions within crowd sequences of real surveillance video. The methods use novel techniques for detecting diverse actions: a motion-speed invariant feature descriptor made from a key-point trajectory, and a weighting and clustering for the trajectory features. We conducted several experiments on the proposed methods, in which our previously proposed single-trajectory method was used as a baseline for comparison and the dataset was that of the TRECVID Surveillance Event Detection task. We discuss how to select the proper method to detect actions in crowd situations through an analysis of these experimental results.
