Robots have played an increasingly important role in various fields of society such as industrial manufacturing, family life, national defense and security, while also placing higher demands on their accuracy, real-time performance and efficiency. Object detection and grasping pose estimation are the basic abilities that industrial robots should possess, playing a crucial role in industrial sorting, palletizing, assembly and other work. Deep learning technology extracts features through neural network, which can greatly improve the detection speed and accuracy under the guarantee of big data and computing ability. In this paper, through the research of target detection and manipulator grasping algorithm based on convolutional neural network (CNN), a multi-target tracking algorithm combining YOLOv3 target detection algorithm and kernel correlation filter (KCF is proposed. The experimental results show that the KCF algorithm has good tracking performance and good fault tolerance for small amounts of occlusion. Thirdly, this paper proposes a visual guidance method for robot grasping combined with Kalman filter prediction, which can guide the robot to accurately complete the grasping task. Finally, this article built a robotic arm sorting system and ported the YOLOv3 network to the Xavier platform. The experimental results show that the network can effectively detect and recognize targets in the experimental environment. The test results verify the effectiveness and availability of the target sorting system.
|