Research On Action Recognition Technology Based On Time Series Analysis

The application of human behavior identification has been extended to almost all aspects of society,especially in sports analysis,public security monitoring,and virtual reality,etc.,which bring great opportunity for many industries’ evolution.Therefore,the study of human behavior recognition technology is the research hotspot in the field of information technology.Although there is a variety of behavior recognition algorithms have been proposed,due to the complexity of human action,the behavior recognition technology still needs to be further developed and upgraded.In this paper,the study mainly focuses on the lack of time series information in human behavior recognition,and implement research on the identification of human motion based on avideo sequence.The main contents are listing as follows:1.A unification of extracted features is carried to overcome the interference caused by the external factors on behavior recognition results.3D skeleton coordinate feature plays a crucial role in behavior cognition,however,the original coordinate data is not robust to it,which leads to the incorrect identification of the effects of scale,angle,and active area.To solve this problem,this paper has carried on the centralization of the coordinate data,the scale,as well as the angle.Therefore,the coordinate feature realizes the invariance of scale and rotation,what’s more,good robustness to the change of the active position.2.Multi-feature fusion is utilized to solve the recognition limitation resulting from a single feature.There are various of skeleton features,the distance between two points,the angle between lines,the angle between line and surface,and lines direction,etc.An excess of features will conduct redundancy while a single feature cannot fully express the movement information.Hence,an efficient behavior recognition requires an appropriate features fusion.In thecurrent paper,a variety of skeletal characteristics is analyzed and also conducted a test to achieve an efficient fusion to better express the characteristics of the movement information.3.Sparse representation algorithm with time series information is applied as the multi-dimensional time series modelling method to overcome the problem of the lack of time information.Although the existing underlying features can express the body action information well,for the movement of thehuman body in the video,the time information is a quite considerable for its distinguishing features,it is because that the video of body movement itself is the changes and movement of frames in the timeline.In order to solve this problem,the regulatory rules of supervision time series chaos are adopted to suppress the error on movement time,which transfer the traditional sparse coding into a multi-dimensional time series modelling method incapable of handling motion information.4.The minimum error matching strategy is used to effectively and simply achieve the correct classified identification of behaviors.There are many methodologies of behaviors classification,such as support vector machine,K-nearest neighbor,etc.Among them,minimum error matching strategy is adopted in this paper for itssimpleimplementation,as well as guaranteeing the successful identification results.Experimental operation of the hardware and software platform provided by the Chinese Academy of Sciences,the experiment carried out in the Linux operating system and Matlab2015 b development platform.The validity of the algorithm is improved by comparing the experiments on the RGB-D behavior recognition database of UTKinect and Florence3 D and the experiment on the artificial simulation data set.