Human Activity Recognition From 3D Skeleton Sequence Based On Bag Of Words Model

In recent years,the challenging work of human body recognition has received more and more attention from researchers.Human activity recognition is widely used in many fields such as smart home,video surveillance,medical health,virtual reality,etc.However,due to the influence of various factors in the human body during activity,such as illumination intensity,background change,and blockage of one's own body,etc.The capture of information and the accuracy of recognition are not ideal.The appearance of Kinect makes it easier to capture depth images of the human body,and the human skeleton information extracted from depth images greatly improves the accuracy of identifying human behavior.This article is based on three-dimensional human skeleton information extracted from depth images.From the aspects of feature extraction and feature coding,the paper takes the information from the underlying information to the high-level characteristics,and finally uses these characteristics to identify human activity.The research results are as follows:1? Based on the moving pose descriptor and hard vector coding method,this paper proposes a human activity recognition based on moving pose description sub-characteristics and improved hard vector coding.The moving pose description sub-frame uses the position information of the 20 joint points and its first and second derivatives to represent the motion information of the human body.In the feature coding phase,because of the traditional hard-vector coding,only the most recent visual word weighted value 1,the rest of the visual word weight is 0.This results in a rough reconstruction of local features,which is not conducive to improving the accuracy of human recognition.An improved against the problems of hard vector coding,namely on the basis of the traditional hard coding for a visual word weighting recently,expand to k visual word weighting,and weighted way is diminishing at from near to far right value.The most recent KTH visual word weight is 1/(k~2+k).The improved coding method can solve the fuzzy problem of visual word more effectively than the traditional hard vector coding,so as to improve the accuracy of recognition.2? Aiming at the problem of low activity recognition rate from 3D human skeleton sequence based on a single descriptor and a single feature coding method,a method is proposed based on multi-descriptor feature coding.The angle descriptor uses the information of 35 joint angles and its first derivative to represent the angle change of human motion.In the process of identifying the body,Firstly,moving pose descriptor and angle descriptor are extracted respectively from 3D human skeleton sequence.Then,vector quantization coding,sparse coding and locality constrained linear coding are employed respectively to get six kinds of feature based on two kinds of descriptor.Finally,linear classifiers are respectively constructed based on these six kinds of feature,and the recognition result is decided by voting strategy.