Human behavior classification from MPEG compressed videos

A new approach is proposed for human behavior classification from MPEG compressed videos. Moving objects are first detected by subtracting the dc values in I frames from those in the background. In addition, the dc values of the background are also adapted to avoid noise and illumination change. The tracking process is then performed using the consecutive frames. Motion vectors extracted from P frames are used to predict the next position of moving objects. An overlapping table is constructed to determine relationships between moving objects, and the number of moving objects is updated. For analyzing human behavior, motion vectors and velocities of moving objects from P and B frames are extracted. These features are clustered to codewords using a codebook generated by vector quantization (VQ) for the input of discrete hidden Markov models (HMMs). By applying the HMM, four kinds of human behaviors are successfully identified from the human behavior sequences. The proposed approach is, furthermore, more accurate than the previous method based on conventional features.