Object Detection Algorithm Based on Lightweight YOLOv4 for UAV

With the development of artificial intelligence technology, the mission scope of UAV has gradually expanded from the traditional image acquisition of high-altitude target area to aerial photography of cities, mapping regions, forest inspection, plant protection monitoring, rescue missions, fire rescue and other aspects. In recent years, the industrial development of UAV has reduced the cost of UAV, improved the performance of camera, stability and other aspects, and more UAV has been put into real life. Among them, the research on object detection in UAV images has become a hot topic, and a large number of excellent algorithms have emerged, aiming at improving the accuracy and speed of object detection. The object detection of UAV is faced with the following problems: small detected targets in the field of vision, weak computing power of the embedded system, weak real-time detection, etc. By comparing some classical object detection algorithms, using YOLOv4 as the basic network model, the structure of the network and the algorithm of image prediction box selection are improved. Then, a lightweight YOLOv4 object detection algorithm for small target is proposed. It can not only improve the speed and accuracy of object detection for the UAV, but also apply to mobile or embedded platform of the UAV. What's more, the realization of real-time object detection greatly improves the UAV's scene understanding ability and operational efficiency.