The beacon spot position detection technology is widely used in the field of vision-based optical communication coarse alignment, and its detection algorithm directly affects the accuracy of acquisition and positioning. Aiming at the defect that the algorithm of searching beacon spot based on threshold segmentation is easy to be affected by background strong light, a real-time acquisition and positioning system of unmanned aerial vehicle optical communication based on deep learning algorithm is established in this paper. First, the YOLOv4 (You only look once, v4) network is improved, four simplified modules and one up-sampling module are designed using the feature map channel splicing method that can enhance the extraction of shallow feature information, which greatly improves the speed of the network. Then, the improved network, the original YOLOv4 network and its simplified network are trained on PASCAL VOC data set. Finally, collect and train the beacon spot data set, and run the improved YOLOv4 network on the unmanned aerial vehicle to output the beacon spot position of the camera video frame. Based on proportion integration differentiation algorithm, the gimbal is adjusted for position closed-loop control, so as to realize real-time acquisition and positioning for optical communication. Experimental results show that the accuracy rate of the improved YOLOv4 network on the beacon spot test set is 99.6%, the recall rate is 99.8%, and the frame rate on the NVIDIA Jetson Xavier NX embedded computer platform is 42 frame/s, which meets the requirements of real-time acquisition and positioning for unmanned aerial vehicle optical communication.