Rotated object detection based on improved Hough transform

Most conventional detection methods are usually with rectangular or square bounding boxes, which may leave out angle information. Therefore, for rotated objects, these detection models become inaccurate. In this paper, based on self-built database, we use mirror image and rotation to expand our data size. Then detect lines and acquire angles of inclination by improved Hough transform. Subsequently, we use an algorithm to filter outliers and estimate direction of objects. According to the aspect ratio of training objects, we utilize rotation angles to predict potential detection boxes and the Non-Maximum Suppression (NMS) algorithm to remove redundant frames for all inclined frames, and get the optimal result. The experimental results show that this method achieve an excellent performance, particularly for symmetric or linear objects and there exists tiny error rate comparing with annual measurement.