Deep learning based time sequence motion identification method

Abstract

The invention discloses a deep learning based time sequence motion identification method. The method comprises video feature extraction and construction of a time boundary regression model. Aimed at the problem that that a long motion feature cannot be expressed effectively enough in the boundary detection process, inter-frame information and intra-frame information are extracted simultaneously via a double flow network to obtain a feature sequence of a video unit, a multi-scale short motion segment extracting scheme combining context information is provided, the accuracy of subsequent regression is improved effectively, the time boundary model is trained via a feature sequence, training time of the model is reduced, and the calculation efficiency is improved. The improved time boundary regression model including an improved multi-task multilayer sensor and a brand new splicing mechanism aimed at long motions is provided aimed at the problem of inaccurate boundary regression of the long motion; and on the basis that the motion type is accurate, the accuracy of time boundary regression of the long motion is improved effectively, the predicted motion segment is more overlapped with the practical motion segment, and the identification rate of the time sequence motions is improved.

Description

technical field [0001] The invention belongs to the field of computer vision and pattern recognition, and relates to a time sequence action recognition method based on deep learning. Background technique [0002] With the rapid development of smart phones and the Internet, video data has begun to show a blowout phenomenon, so research in the field of computer vision is also gradually expanding in the direction of video data. The basis of video processing is action recognition. Although traditional action recognition has achieved a high recognition rate, because the original data must be a short video with a fixed frame number after cropping, it is required to include a single action tag. Such cropping requirements are too harsh. , but in practical applications, actions appear randomly in long videos, so traditional action recognition algorithms cannot meet the actual application scenarios. Time-sequence motion detection is a specific study of such uncropped original long vi...

AI Technical Summary

Problems solved by technology

On the basis of accurate detection of action categories, in order to solve the problems of insufficient expression of long action features and inaccurate regression of long action boundaries in the proc...

Images