Research On Human Behavior Recognition Based On Kinect And Bag-of-words Model

Human behavior recognition is one of the hot research directions of computer vision.The emergence of low-cost depth sensors,such as Microsoft Kinect,can effectively avoid the influence of light,environment and other factors.It can obtain the depth information and skeleton information in the image,which provides a good data source for human behavior recognition research.Therefore,in the research of human behavior recognition,researchers gradually tend to use three-dimensional skeleton information.As a machine learning method,Bag of words(BoW)model is widely used in human motion recognition based on skeleton joints.At present,the recognition accuracy of this method still needs to be further improved.According to the current research situation,this paper mainly carries out the following work:(1)The recognition accuracy can be influenced by the shape of tester,as well as the distance and angle between tester and depth sensors in the Kinect based human action recognition.Therefore,this paper proposes the adaptive adjustment method and further to process uniformly the rotation angle of the skeleton motion,which can effectively process the depth datasets with different joint number and camera angle.(2)Recently,the methods for constructing feature vectors can be divided into three categories: coordinate,distance and orientation.The distance representation is robust to the position and orientation of human body relative to the camera.The orientation representation is robust to position and human body,shape of human body and the orientation relative to the camera.Based on the above three representations,this paper proposes to construct the action feature vectors by combining spatial descriptors and temporal descriptors,in which the spatial descriptors are constructed by the coordinates of joints,the angle between skeleton vectors,the direction angle and elevation angle of skeleton vectors and the temporal descriptors are constructed the difference of joint coordinates between frames.(3)In the traditional BoW model,the visual vocabulary is constructed with Kmeans algorithm whose performance is subject to the initial cluster centers.This paper improves the visual vocabulary construction method and proposes to construct the candidates set for the data points with larger hub values by using hubness phenomenon of high-dimensional data set.And the visual vocabulary is determined by maximin method.We evaluate the effectiveness of the proposed method on five public datasets,including CAD-60,UTKinect,MSR action 3D,UTD-MHAD and MSRC-12 datasets.The experimental results show that the proposed method performs better than several state-of-the-art algorithms.