Abstract: Forest fire has the characteristics of uncontrollability and random spreading, so forest fire prevention is very important. Due to the vast forest area, the traditional forest fire monitoring method has many disadvantages, such as blind area, poor real- time performance, large resource consumption and so on. In this paper, a forest fire monitoring system is designed and implemented based on the target detection technology with UAV as the flight platform. Aiming at the poor performance of the original ecological YOLOv3 algorithm in smoke and flame detection, the algorithm is improved by adding NMS and re clustering anchors. The detection accuracy and recall rate are increased by 6.8% and 11.92% respectively. The system can monitor forest fire in real time, and improve the ability of forest fire prevention and the automation and digitization level of forest fire warning.

Key words:  forest fire, UAV, target recognition, YOLOv3