Research On Human Behavior Recognition Based On Bone Features

With the development of science and technology such as artificial intelligence,chip manufacturing,and intelligent control,robots with certain autonomous intelligence characteristics are widely used in production,life,military operations and other scenarios.However,because robots do not have complete autonomy,they cannot be completely independent.Therefore,one of the current research directions is to study human-machine collaboration to complete tasks in complex scenarios by combining the characteristics of human’s analysis and understanding of dynamic tasks in complex environments and autonomous decision-making.Human-robot collaboration refers to the cooperation between humans and robots to complete specified tasks.At present,industrial robots are widely used in manufacturing to complete product assembly work.It enables human beings to complete more work related to design and creation.In human-robot collaboration,how robots understand human intentions is an important research direction in human-robot collaboration technology.From a traditional point of view,humans rely on joysticks and command lines to instruct robots to complete tasks.This manipulation method makes the robot less flexible.At the same time,the requirements for operators are relatively high,and various technical instructions need to be mastered.With the development of artificial intelligence technology,there are many ways for robots to understand human intentions,including speech recognition,gesture recognition,behavior recognition and so on.In the existing research work,most scholars will focus on behavior recognition,classify and recognize human behaviors in video streams,in order to infer human intentions and complete human-machine collaboration.This recognition method requires lighting,venue,etc.And,if the video contains more noise,the recognition accuracy is not high.Therefore,in order to further study the behavior recognition technology,this paper considers the classification and recognition of human behavior based on human skeleton data,so as to realize the judgment of human intention and complete the human-machine collaboration tasks.This paper takes the human-robot collaborative assembly task as an example,and establishes a human-robot collaborative assembly scene in the laboratory.The specific task is to assemble a chair through human-robot cooperation.The human is the leader in the assembly task,and the robot acts as a facilitator to deliver chair accessories(such as chair legs)or assembly tools(such as an allen wrench)to humans.The robot can discern human behavior and actions,and predicts the next assembly intention of human beings and prepares accessories or tools in advance.The research work done in this paper includes the following aspects:(1)This paper first analyzes the theoretical basis of human action recognition technology,including classic supervised learning algorithms such as support vector machines and backpropagation neural network,and points out the shortcomings and deficiencies of these algorithms.the basic theory of convolutional neural network and recurrent neural network are introduced,and the selection strategy of behavior recognition algorithm based on deep learning is analyzed,and then the behavior recognition scheme based on convolutional neural network is determined in this paper.(2)Then the software and hardware system of the Kinect sensor is analyzed,and the external parameter calibration of the Kinect sensor is completed by using the three-sided target reconstruction technology.The extrinsic parameter matrix of the two Kinect sensors is calculated,and the human skeleton data in the global coordinate system is obtained,and then the error analysis is carried out on the data collected by the Kinect sensor.(3)Based on the fact that two Kinect sensors can effectively collect human skeleton data,this paper further designs a human-robot collaborative assembly scene,and divides the process of human-robot collaborative chair assembly into seven steps.The robot should predict the next behavior of human beings accurately and prepare accessories or tools in advance.In this paper,the Kinect sensor is used to collect human behavior and motion data,and the data preprocessing is completed.(4)Considering the lack of feature understanding of the convolutional neural network,a dual-stream neural network was introduced,and a behavior recognition network model based on the attention mechanism was further designed.The model was used to classify and recognize human skeleton data.The results show that compared with methods such as CNN and LSTM,the improved two-stream neural network proposed in this paper has a higher accurate recognition rate,which shows the effectiveness of the algorithm designed in this paper.