Hand-eye System Research Based On YOLO-V3 Target Detection Algorithm

In recent years,with the progress of science and technology,the research of robot hand eye system has become an important direction of artificial intelligence research,and hand eye system is also the most important part of robot system.Before the vision system is given to the robotic arm,the movement track of the robotic arm is usually the planned point-to-point movement in advance.When it is unable to accept the external information,it is unable to make corresponding adjustment according to the environmental changes.The combination of the vision recognition system and the robotic arm grabbing system solves this problem.The use of the hand eye system improves the intelligent degree of the robotic arm,which is of great importance Research significance and application value.This thesis mainly studies the hand eye system based on the object detection algorithm of YOLO-V3.The main work flow is as follows:firstly,the vision sensor collects the image information,obtains the image coordinate information of the object through the image processing and detection positioning,then obtains the 3D coordinate information of the object through the camera internal and external parameters obtained by the calibration method,and finally transmits the 3D coordinate information of the object The robot arm grabbing system completes the grabbing task.The main work can be divided into the following parts:the first part is the target detection algorithm,the second part is the YOLO-V3 target detection algorithm experiment and improvement scheme,the third part is the hand eye calibration experiment based on Kinect-V2 vision sensor,the fourth part is the hand eye system grabbing experiment.In the vision system of this thesis,as a more advanced algorithm in the field of Deep Learning,YOLO-V3 algorithm is not only faster in detection and recognition,but also higher in detection accuracy compared with other deep learning algorithms.In order to solve the problem of missing detection and false detection caused by partial occlusion,an improved scheme based on Soft-NMS is proposed,which greatly improves the detection speed and accuracy.After obtaining the pixel coordinates,the pixel coordinates need to be transformed into the three-dimensional coordinates under the Dobot robotic arm base coordinates.In this thesis,we use Zhang Zhengyou calibration method to calibrate Kinect-V2 camera and hand eye calibration experiment,obtain the camera internal and external parameters,and give several methods to improve the calibration accuracy,and realize the transformation of image coordinates to three-dimensional coordinates.In the last part of the hand eye system grabbing experiment,the whole hand eye system is built based on the Deep Learning Machine,Kinect-V2 vision sensor and Dobot robotic arm,build the whole software system based on ROS.The experiment of target recognition,location and grabbing is carried out to verify the success rate of the system's independent grabbing and error analysis.The result shows that the hand eye system based on theYOLO-V3 target detection algorithm can successfully grab the target,it has feasibility and practical application value.