Unmanned aerial vehicle (UAV) object detection algorithm based on keypoints representation and rotated distance-IoU loss

Recently, significant progress has been made in the research field of unmanned aerial vehicle (UAV) object detection through deep learning. The proliferation of unmanned aerial vehicles has notably facilitated the acquisition of corresponding data. However, the presence of substantial rotated objects in various orientations within UAV data sets poses challenges for traditional horizontal box object detection methods. These conventional approaches struggle to precisely locate rotated objects. Consequently, algorithms for rotated bounding-box object detection have been proposed; however, some of these existing methods exhibit issues, including periodicity of angle and exchangeability of edges. We propose a joint key point representation and rotated distance loss object detection network to solve the above problems. It is mainly composed of the key point representation module and the rotated distance-IoU loss. The key-points representation is used to indirectly represent the angle parameter of the rotated bounding box. It accomplishes this by measuring the angle between the line connecting the center point of the rotated bounding box to a specific boundary center point and the horizontal line. Next, the coordinates of the center points of anchor and the center points of its boundary are used to obtain the height dimension of the rotated bounding box and the width dimension of a rotated bounding box is introduced. Like this, the rotated bounding box can be represented by two points and a width dimension. Also, based on the traditional rotated IoU loss which does not incorporate the distance between the center point of the prediction box and the center point of ground truth in the regression process, the rotated distance-IoU loss is proposed to replace the traditional rotated IoU loss, which speeds up the convergence of the network. We have conducted extensive experiments on the DOTA data set and the DroneVehicle data set and have demonstrated the effectiveness of the proposed method.