Human Activity Recognition Of The Multi-channel Feature Combination Based On Kinect

The ability to recognize hunan actions is one of the most popular research topics in computer vision and artificial intelligence,in the video surveillance,human-computer interface,sports video analysis,video retrieval,robotics and other research areas have also been widely concerned.The traditional research on activity recognition generally focus on RGB,with the advent of depth sensors,more and more scholars study human action recognition based on depth images.Compared to RGB,depth images have robustness to changes in lighting,color,and texture.Additionally,the location of the skeletal joints can indicate the relative position of human skeleton.Based on the existing methods of human action recognition,this paper presents a research method that is human activity recognition of the multi-channel feature combination based on Kinect.In the first,we collect the three channels of data information with Kinect,which contains RGB,depth and joints of the skeleton,and then extract the feature information of these three channels by different feature extraction methods.For RGB,spatial interest points are extracted and the histogram of oriented gradient(HOG)features for each interest point from both image and motion channels are computed,for depth,a new descriptor for spatio-temporal feature are obtained by applying a modified HOG and for a human object,20 joint positions are tracked by the skeleton tracker Kinect,in which,for each joint,we extract the pairwise relative position features,and then use the Fourier-temporal-pyramid to represent the temporal dynamics of these features,after then,coding and pooling the low-level features.We apply bag-of-words to code these features,and transform the encoded features into histogram to represent the video,in the next,normalize the histogram to obtain the feature representation of the entire video.Finally,a cascade SVM was used to establish a training model for human actions recognition.In this paper,the proposed algorithm is trained and tested on four standard datasets,and compared with the state-of-the-art results,experimental results show that the multi-channel feature extraction algorithm used in this paper has good performance.Finally,the Kinect sensor is used to collect the data of the real world to verify the practical application of the proposed algorithm.The results show that the proposed algorithm has better recognition accuracy in real dataset.