Video-based behavior recognition method

Abstract

The invention discloses a video-based behavior recognition method, and belongs to the field of video image processing. The video-based behavior recognition method comprises the steps of converting to-be-detected video data into an RGB frame and an optical flow frame, putting the RGB frame and the optical flow frame into a trained sub-network to obtain feature values of the RGB frame and the optical flow frame, and putting the feature values into a trained long-short-term memory network to obtain a behavior recognition result, wherein the child network is supervised by the parent network trained through cross fusion during training. According to the video-based behavior recognition method, the accuracy of behavior recognition is further improved by using cross fusion, and the problems thata traditional algorithm is low in accuracy and long-time-period information cannot be effectively utilized are properly solved.

Description

technical field [0001] The invention relates to a video-based behavior recognition method, which belongs to the field of video image processing. Background technique [0002] With the continuous development of behavior recognition technology, video-based behavior recognition becomes more and more reliable. Compared with using still images for classification, video image information can provide an additional important clue: time component. Using the temporal body movement information of the actor in the video can identify many actions more reliably, and then classify the video. In addition, video provides natural data augmentation (dithering) for single still image (every frame of video) classification. [0003] Video classification and behavior recognition have attracted great attention in the academic circles due to their wide application in many fields such as public security and behavior analysis. In action recognition, there are two key and complementary aspects: appe...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to: aim at the defects that the common dual-stream CNN only combines RGB-based CNN and optical flow-based CNN for behavior recognition at the end, and cannot reasonably integrate the advantages of optical flow images and ordinary RGB images, and the effective data is not sufficient problem, a video-based behavior recognition method is proposed, which effectively solves the problems that ordinary dual-stream CNN cannot effectively use time information and insufficient data volume, and the recognition accuracy is significantly improved.

Images