When unmanned aerial vehicle (UAV) performs fire extinguishing task, the spraying direction of nozzle is fixed and the direction of fire point cannot be found automatically. By manually controlling the height and pitch angle of UAV, the spraying direction is tried to be adjusted continuously to ensure that extinguishing agent is sprayed on the flame to realize effective fire extinguishing. Aiming at the low efficiency and low accuracy of manual fire point detection and directional injection, a method is proposed to realize real-time target detection in low-cost and low-power hardware, and then control the directional fire point of nozzle. The trained YOLOv2-Tiny target detection model is deployed to the mobile platform K210, and the internal convolution neural network hardware accelerator quickly detects whether there is a fire point in the video and outputs the position information of the target. According to the position information, PID control algorithm is used to adjust the directional fire point of the two-dimensional rotating nozzle. The model is trained by using the data set constructed by surveillance camera shooting and downloading on the Internet. After training, the detection accuracy of the model reaches 94.60%. When the model is downloaded to K210 development board, the speed of 14 frames per second can be reached, the detection accuracy reaches 96.5%, and the average error of landing position is 3.7 cm, which has high practical value.