Abstract: Aiming at the limited environment of PC (Personal Computer) in the application of UAV fire detection, a method of UAV fire detection and PTZ tracking based on YOLOv2-Tiny is proposed. First, perform pre-training on the improved YOLOv2-Tiny model to obtain the optimal YOLOv2-Tiny model, and deploy the optimal YOLOv2-Tiny model on the K210 development board. Secondly, transmit the detected fire image to the cloud and transfer the fire frame. The distance parameter between the selection center and the image center is passed to the PID process to control the pan-tilt to realize real-time fire tracking. Finally, the ability of fire detection and pan-tilt tracking is verified through the actual flight of the drone. The experimental results show that compared with the YOLOv2 model, YOLOv2-Tiny has a higher detection rate on the test set, the detection rate reaches 96.66%, and the detection speed reaches 14 frame persecond. The PTZ tracking center position pixel error (CPE) is lower than 5, and the UAV attitude angle remains relatively stable during real-time detection and tracking. This research has potential in real-time fire detection.

Key words: UAV; YOLOv2-Tiny; PTZ tracking; K210 development board; PID