Abstract: In order to accurately identify the complete action sequence of video target individuals and enhance the accuracy of video behavior recognition, a deep network time grouping behavior recognition method based on over-limit learning machine is proposed. First, the number of states of the behavior recognition model is delermined according to the number of key human behavior gestures, establish the multi-scale structure association of human motion behavior, and the different scale characteristics of motion trajectories and edge contour wavelet moments are introduced into the behavior model to obtain general characteristics of human motion behavior. Using the video grouping sparse sampling method, the video is divided into equal duration groups, and the standard backpropagation method is used to optimize the model parameters, to realize low-cost video-level time modeling and to ensure the integrity of the modeling process information. Finally according to the hidden layer activation function output and corresponding output layer weight coefficients, sensitivity analytical formula is obtained, hidden nodes are sorted according to sensitivity parameters, minor nodes are deleted, and accurate recognition of deep network time grouping behavior are realized. The results of simulation experiments show that the method has a high level of recognition accuracy, is less time-consuming, and has strong robustness.

Key words: regularization, overrun learning machine, deep network, behavior recognition