UAV are widely used in rescue, logistics, inspection and other scenarios, and timely detection of obstacles during flight is a prerequisite for safe completion of tasks. In order to meet the requirements of obstacle detection in UAV flight, an improved algorithm was proposed based on YOLOv5 algorithm. The convolution module was used instead of the Focus module and a detection probe was added to the original network structure to improve the detection performance. The CA (coordinate attention) attention mechanism was integrated in the feature extraction network to strengthen the feature extraction ability of the model and reduce background interference. Experimental results show that compared with the original YOLOv5 model, the improved algorithm improves the accuracy by 13.8%, the recall rate by 12%, the mAP@0.5 by 15.4%, and the mAP@0.5:0.95 by 10.2%, which proves the superiority of the improved algorithm.