Aiming at the key technologies of visual collision avoidance in UAV obstacle avoidance system the YOLOv4-tiny target detection algorithm is improvedand a visual collision avoidance method based on target detection is proposed.With the separable convolution structure of MobileNetthe YOLOv4-tiny target detection network is optimized to improve the detection accuracy.The efficiency of the SURF matching algorithm is improved by removing the redundant information of the undetected target area.The target area is calculated in three dimensions by least squares methodand the calculated depth and pose information of UAV are treated as the basis of visual field scale of the collision avoidance determination areathe visual collision avoidance of UAV is realized through the integration of the pixel position of target detection.Finally the improved network model is analyzed by using standard data setthe mAP value is 69%and the UAV visual communication module is built for visual collision avoidance testwhich verifies the rationality of the method.