Human Behavior Recognition Based on Wavelet Moment and Regional Optical Flow

Motion history image (MHI) and motion energy image(MEI) can recognize simple actions effectively, but they can't represent the velocity information of the actions. Sometimes in the behavior recognition, we must take the speed into account. For example, touching a person, fierce and friendly touches are two totally different movements. The paper combines the wavelet moment of temporal motion descriptors (MHI and MEI) and the speed feature based on optical flow to represent the action. Wavelet moments are rotation, translation and scale invariance. Foreground is obtained from video sequences by background subtraction. Then we use Lucas-Kanade algorithm to get the regional optical flow information, whose direction amplitude can represent the velocity changes in different direction intervals. Experiments based on video sequences outdoor scenarios carried out to verify the effectiveness of the proposed method.