Research On Robot Human-robot Interaction System Based On Deep Learning

In recent years,driven by the "Industry 4.0 Revolution" and artificial intelligence technology,robot technology has been developing rapidly.Whether in the industrial field,life field or military defense field,robot technology has a wide range of applications,and has made a series of major breakthroughs.As the main research object of robot technology,robots,especially intelligent robots,play a pivotal role in people’s life and work,and people also put forward higher requirements for the way of human-robot interaction.However,the current traditional human-robot interaction methods based on keyboard and mouse,teaching device and wearable devices often have problems of high cost,poor real-time performance and low efficiency.Therefore,in order to build a natural and efficient human-robot interaction interface,this paper combines artificial intelligence technology and robot technology to build a robot human-robot interaction system based on deep learning.In this system,the robot can automatically locate the target object according to the operator’s gesture instruction,and then finish the operation of grasping and placing the target object autonomously.The main work completed in this paper is as follows:(1)In order to provide a new and intelligent way of human-robot interaction,this paper built a human-robot interaction system combining gesture recognition module,object detection module and robot positioning and grasping module,and completed the internal and external parameter calibration experiment of Kinect camera and the forward and inverse kinematics solution of EPSON industrial robot.(2)Due to the fact that the original YOLOv7 network model is prone to misdetect and miss small target gestures in complex environments,CBAM attention module and FRe LU activation function are introduced on the basis of the original network,and the gesture recognition algorithm based on the improved YOLOv7 network is constructed.A gesture instruction data set was made,and the training and evaluation of the improved YOLOv7 network model was completed on this data set.The results show that the improved YOLOv7 network has a very good recognition effect on small target gestures in complex environments.On the self-made data set,the m AP@0.5 and m AP@0.5:0.95 of the network reach 99.6% and 79.2%,respectively,which are improved by different degrees compared with the original network.(3)For the problem that Center Net network cannot detect the rotation angle of objects in human-robot interaction scenes,this paper adds the rotation angle regression branch on the basis of the original network,selects DLANet as the backbone network for feature extraction,and replaces the L1 loss function used in the original network with Smooth L1 loss function.Therefore,R-Center Net rotating target detection network is built.By using this network,the information of object category,center pixel coordinate and rotation angle can be obtained.A robot rotating target data set was developed and the R-Center Net network was trained and evaluated on the data set.The results show that the rotating target detection network has high recognition and positioning accuracy.(4)The performance of the human-robot interaction system is verified through experiments.Firstly,the human-robot interaction experiment platform was built,and the software design of the human-robot interaction system was completed based on the TCP/IP protocol.Then,the gesture recognition effect of the improved YOLOv7 network and the rotating target detection effect of the R-Center Net network were tested in the actual scene.Finally,the gesture interaction grasping experiment was carried out on the experimental platform built,and the experiment verified the stability and efficiency of the robot human-robot interaction system based on deep learning in this paper.