I. Introduction
The research of machine vision appeared in the 1950s. Compared with other traditional fields, the history of machine vision is short, and the development speed is limited by the development of computers. It wasn't until the late 1970s, when computer performance was capable enough to handle large-scale data such as images, that computer vision gained serious attention. The rapid development of machine vision has greatly accelerated the intellectualization of traditional industries. The pan-tilt system combining mechanical structure, machine vision and control circuits is widely used in robotics, industrial automation, monitoring, military and other fields. The structure of the pan-tilt is mainly based on the linkage mechanism nowadays, the camera or other actuators can be placed at the end of the system, the positioning accuracy is high but the physical interference between the links will limit its range of motion, and the overall structure is not compact enough. On the other hand, the real-time of the vision algorithm for embedded platforms needs to be improved. In this paper, a scheme of target recognition and tracking of the ball gear pan-tilt system based on YOLOv4-tiny and DeepSORT algorithm is designed combining with PID controller. The system can realize the real-time tracking of the target figure according to the experimental results.