A violence recognition method based on cross-frame timing feature enhancement

Abstract

The present application belongs to the technical field of computer vision, video understanding and intelligent security, and particularly relates to a violent behavior recognition method based on cross-frame time sequence feature enhancement, wherein a video input and preprocessing module forms a space-time Token sequence after frame sampling and preprocessing; the space-time Token sequence is encoded by a video Transformer backbone feature extraction module to obtain a basic space-time representation; a cross-frame motion difference enhancement module enhances the motion difference of adjacent time segments in the basic space-time representation space; an adaptive frame-level time sequence attention module performs residual enhancement on the time segments according to the frame-level importance; a mean-max dual-path hybrid pooling module fuses the global average semantics and the local maximum response; and a classification output module outputs the violent behavior recognition result according to the fused video-level features. The present application solves the problem of insufficient modeling of local motion changes, key time segments and local strong response regions in the existing method, and improves the expression ability and recognition ability of the model for the space-time features of violent behaviors in monitoring videos.

Images