Human Behavior Representation And Recognition Method Based On Kinect

With the rapid development of information technology,the disadvantages of using traditional two-dimensional image information to identify human behavior become more apparent.Recognition under three-dimensional scene in real space has gradually become the dominant direction in the field of human behavior recognition.Kinect somatosensory devices have extensive applications in areas such as virtual reality,smart home,and behavior analysis because of their excellent depth perception and human body 3D joint recognition capabilities.With the rise of big data and artificial intelligence technologies,traditional behavior recognition algorithms that focus on deep tuning of behavioral models have problems such as insufficient data storage space,low recognition efficiency,and poor scalability.Spark is a elastic distributed data set model based on memory,which can significantly improve the efficiency of data computing,especially for the data intensive algorithm of iterative operation.Therefore,this paper studies how to use the information of human body joint points in the space obtained by Kinect to represent human behavior,and on this basis,we combine the Spark MLlib algorithm library to explore the human behavior recognition.For the effective development of research work about the behavioral representation and identification method,a complete experiment system platform was constructed using Kinect as the input medium for joint point information,using the distributed file system HDFS as the data storage platform,and using Spark framework based on in-memory computing as a basis for computing.Through experiments,the feasibility of the original joint point data,angle features and distance features in expressing human behavior was analyzed.The static behavior representation method was constructed based on the angle features.The sequence of angle features between static frames during the process of dynamic behavior was analyzed,and use the key frames in the dynamic behavior process are obtained by the dynamic behavior key frame extraction method using cosine similarity,and the dynamic behavior representation method based on key frame angle feature change sequence is proposed.Self-built behavioral data set combines random forest algorithm of Spark MLlib algorithm library to carry out experiments for the behavior recognition,the experimental results show that the accuracy of the algorithm for 10 static behaviorscan attain 99.7% and the accuracy of 10 kinds of dynamic behaviors can attain 91.8%.In order to improve the generalization ability of the random forest behavior recognition model,this paper uses the ensemble learning idea of Random Forest,a multiple random forests algorithm with weighted large number voting is proposed by using uses the parallel and rapid iterative features of the algorithm under the Spark platform.The experimental results show that the accuracy of dynamic behavior classification have obviously increase with the increase of the number of base classifiers.The behavior recognition accuracy can tend to be stable and can attain over 95% when the number of base classifiers is over 5.The utilization of MSR Daily3 D data set verifies the effectiveness of the behavior recognition and classification method in this paper.