To improve the accuracy of obstacle localization, a data-augmentation-based monocular visual obstacle localization method for UAV is proposed in this paper. The method can be divided into two parts: obstacle detection and obstacle localization. In the first part, the obstacle detection algorithm based on the YOLOv5 network is trained. To solve the problem of difficulty in obtaining samples, the data augmentation method is used to simulate different flight environments and camera equipment of UAV. To enhance the accuracy of detection, add penalty terms for the distance between the detection box and the ground truth box center, as well as the aspect ratio of the two boxes, to the loss function of the detection algorithm. In obstacle localization, according to the conversion formula from pixel coordinates to ground coordinates, the location of the obstacle can be achieved with calibrated intrinsic parameters of the camera and the UAV’s attitude. Consequently, UAV can recognize and locate obstacles and take action to avoid collision.