Human motion identification method based on dense sampling of motion boundary and motion gradient histogram

Abstract

The invention discloses a human motion identification method based on dense sampling of the motion boundary and a motion gradient histogram. The method mainly comprises steps that 1), a video stream is inputted; 2), the optical flow field of the inputted video is calculated, feature point sampling is carried out, and dense feature points are extracted; 3), the trajectory of the feature points is calculated; 4), dense descriptors along the feature point trajectory are calculated; 5), two adjacent frames of video images are derived in time to obtain time series motion images, and the spatial gradient of the motion images is calculated to obtain a motion gradient descriptor HMG; 6), feature encoding is performed separately for each descriptor; 7), after regularization of each descriptor, thedense descriptors and the motion gradient descriptor are connected in series to form a feature vector; 8), the feature vector is trained and learned to obtain a human motion identification model; and9), the human body motion is identified through utilizing the human motion identification model. The method is advantaged in that motion identification precision is improved, and calculation cost is further reduced.

Description

Technical field [0001] The invention relates to the field of machine vision, in particular to a human action recognition method based on dense sampling of motion boundaries and a motion gradient histogram. Background technique [0002] Human action recognition is one of the important branches of computer vision research, and it has great application value in the fields of video surveillance, video retrieval, human-computer interaction, virtual reality, and mobile analysis. However, human action itself has a large degree of freedom. The differences between classes caused by camera movement and viewing angle changes, and the complex relationship between action recognition and human posture, related targets and scenes, etc. bring great challenges to human action recognition. [0003] In action recognition, the feature representation based on the underlying pixels that is usually used has strong robustness to complex backgrounds, but there are limitations in using global or local featu...

AI Technical Summary

Problems solved by technology

Due to the presence of camera movement in the video, and dense sampling will generate too many feature points to be tracked, only some of them can be used for effective feature calculation, resulting in inaccurate representation of human motion features, which also brings a lot of calculations

Images