Research On Action Recognition Technology Catering To Human-computer Interaction

"Artificial intelligence,opening infinite possibility",with the arrival of big data and cloud computing era,artificial intelligence and human-computer interaction technology have increasingly attracted the attention of the scientific community.The interaction between human beings and robots has become a hot topic in the scientific field.Generally,complicated procedures are often used to control the robots.On the contrary,it is simple and convenient to control robots by using human actions,especially for children and the old.Traditional human action recognition based on vision is easily interfered by environmental light,has poor robustness,slow recognition speed,not meeting the demand of real-time interaction with the robot.This paper proposed a new fast human action recognition method,a real-time action recognition and robot control system is designed by using the Kinect camera of Microsoft as the input device.In view of the complexity of environment background and the requirement of rapid action recognition,this paper proposed a method of fast action recognition based on the joint feature.20 joints of human body were obtained by Kinect and from which the key joints were chosen for feature extraction.In this paper,two feature extraction methods were proposed: one is to describe the action by vector sequences of eight fixed key joints;the other is to select the key joints and calculate the angle features from the 16 joints by using the theory of information entropy and select the key joints pairs by the quantitative relationship between the variance of intra-class and inter-class for the calculation of the distance characteristics.A Fast Dynamic Time Warping(FDTW,Fast Dynamic Time warping)algorithm was proposed to identify the actions.Firstly,the lower bound function was introduced to pre-classify the actions and filter out most of the non-target samples.Secondly,in the secondary matching stage,the halfway truncation technique was proposed to terminate the actual distance calculation when calculating the effective path distance,so as to improve the recognition speed.The database containing 20 kinds of actions was set up by the laboratory members.Robustness,recognition rate and recognition time experiments were carried on and the results show the effectiveness of the feature extraction and action recognition methods,realizing the effective control of the robot.