Aggressive motion detection based on normalised Radon transform and online AdaBoost

A framework for human aggressive motion detection in image sequences captured from a single stationary camera is described. Background subtraction, connected chips linking and mean-shift estimation are used to extract target contour information. The binary rectangle containing contour points of the blob-set is transformed into a normalised Radon matrix. With labelled (normal or aggressive motion) Radon matrix as feature pools an online AdaBoost feature selection is implemented. Using the selected classifiers the human aggressive motion in a frame can be recognised. Experimental results show that the system performs well.