Approximated acceleration modes-based detection method for violent video

Abstract

This paper first introduces the Pixel Tracking methods to solve the optical flow information from the original video frames, based on which the optical flow field is established. In this paper, the LFN2-2DCNN-ConvLSTM velocity recognition model is constructed based on the optical flow field, and the Acc-2DCNN-ConvLSTM acceleration recognition model is further constructed based on the velocity features. Among them, this paper innovatively gives a regionalized approximation method to approximate the acceleration feature sequence, obtains an efficient approximation of optical flow feature information under acceleration modes, and obtains the velocity and acceleration modes under CBAM mechanism centralized data. We verified the accuracy of the data in our experiments.