Behavior Recognition Based on Complex Linear Dynamic Systems

Time dynamics is a very important part of human behavior recognition. The linear dynamic system can model the time dynamics, but in the traditional linear dynamic system, the transfer matrix and the output matrix are subject to permutations, rotations, and linear combinations. Therefore, each row in the output matrix can not uniquely identify the characteristics of the corresponding system. In this paper, we propose complex linear dynamic systems to extract the "invariant" features of each time series. Firstly, describing the original video using motion boundary histogram (MBH). Then, we propose to model the motion dynamics with complex linear dynamical systems (CLDS) and use the model parameters as motion descriptors. Finally, the KNN classifier is used to classify it. Experiments with the KTH and UCF sports database show that our method is more accurate than the traditional linear dynamic system.