Research On Kinect Camera Based Human Activity Recognition

Human activity recognition as a hot topic in the field of computer vision,it gets more and more attentions from the researchers and is widely utilized in computer vision community,for instance:smart surveillance,intelligent security,virtual reality,athletic performance analysis and advanced human-computer interaction and so on.Most of existed human activity recognition algorithms calculate low-level or high-level visual features from RGB videos to represent human actions,the recognition rate of these algorithms falls sharply under the influence of noise,clutter background,partial occlusion and viewpoint variations.Thus,extraction high efficiency and discriminative visual features is the key step in the successful human actions recognition system.In recent years,with the emergence of low-cost and simple-operator visual sensors(Kinect),a lot of researchers propose to apply depth and skeleton data which are collected from Kinect camera for human activity recognition.Motivated by this work,This paper focused on human actions recognition against depth and skeleton data.The innovations and contributions are as follows:In order to improve recognition rate and reduce the computational complexity.This paper presents a key skeleton joints based human activity algorithms.A method which combines an improved K-means algorithm with joint movement volume is employed to extract key skeleton joints subset,then further computer depth occupancy pattern(DOP)of the key skeleton joints and 3D key skeleton joint position features to describe human actions.In order to decrease the key skeleton joints features,we utilize Fourier Temporal Pyramid(FTP)to encode these features.Consequently,an L2-regularized collaborative representation classifier(L2-CRC)is applied to recognize human actions.According to the unique properties of depth sequence and the significance of feature fusion in actions recognition.This paper proposes a human activity recognition algorithm which fusing a variety of heterogeneous features.We directly extract a sequence of 3D point cloud from depth sequence,then the multi-scale local spatial-temporal feature descriptor is calculated by a coarse to fine algorithm from 3D point cloud sequence,which is constituted by the multi-scale local position pattern(MSLPP)and the movement distance of two adjacent MSLPPs.In addition,we apply the method of feature fusion to concatenate multi-scale local spatial-temporal feature and key skeleton joints features.Finally,a random decision forest(RDF)classifier is applied to mine discriminative features and recognize human actions.The proposed human actions recognition algorithm in this paper has been evaluated on three benchmark activity datasets.Experimental results demonstrate that the proposed algorithm outperforms the state-of-the-art methods,especially suitable for recognizing interactions which include human subjects and objects.