Human motion recognition feature expression method based on depth mapping

Abstract

The invention relates to the technical field of pattern recognition and specifically to a human motion recognition feature expression method based on depth mapping. The human motion recognition feature expression method based on depth mapping is easy to implement, increases the accuracy of motion recognition, and comprises steps of: extracting, from a color image sequence of a motion-sensing video camera, a human motion space-time interest point p; computing a light stream feature and a gradient feature in the color image sequence by revolving around the human motion space-time interest point p; finding out, from a depth image sequence of the motion-sensing video camera, a corresponding human motion space-time interest point p' and acquiring the depth value of the p'; dividing the p into N layers according to the depth value of the p' and on the basis of a Gaussian mixture model, constructing multichannel expression based on depth mapping, clustering the space-time interest point of each layer by using a clustering algorithm and expressing the space-time interest point of each layer by using a bag-of-word model to obtain a histogram vector; and connecting the feature expression of each layer to form a feature S=(H1,..,Hi,..,Hn). The method is mainly used in the aspect of human motion recognition.

Description

technical field [0001] The present invention relates to the technical field of pattern recognition, and more specifically, to a feature expression method for human body activity recognition based on depth mapping. technical background [0002] Human activity understanding has been widely used in sports training, rehabilitation engineering, ergonomics, game and animation production, security monitoring, etc. The core technology of human activity understanding is mainly target recognition algorithm. So far, activity recognition algorithms based on 2D color information and 3D depth information have achieved some success. However, algorithms based on 2D color information are vulnerable to challenges such as illumination changes and complex backgrounds, and algorithms based on 3D depth information are limited by depth image noise and the accuracy of bone point extraction. The invention patent with the publication number CN104091167A and the name of the feature extraction method...

AI Technical Summary

Problems solved by technology

However, algorithms based on 2D color information are vulnerable to challenges such as illumination changes and complex backgrounds, and algorithms based on 3D depth information are limited by depth map noise and the accuracy of bone point extraction.

The invention patent with the publication number CN104091167A and the name of the feature extraction method of human activity recognition based on the somatosensory camera discloses the feature extraction method using the somatosensory camera, but its LDP generation is relatively complicated and the amount of calculation is relatively large. Therefore, it is necessary to carry out Improve

Images