Abstract:

Existing object detection methods have several problems when detecting traffic objects in uav aerial images， including the poor fittness between the horizontal bounding box and the rotated objects， incorrect suppression due to the high overlap between the bounding boxes， and the mismatch between the sampling points of the standard 2d convolution and the rotated objects.To solve these problems， a single-stage rotated object detection network called ATB-YOLO based on YOLOv3 was proposed. For the backbone of the network， a new feature extraction network called Darknet-53-Dense was designed. The Mish activation function was used to replace the Leaky ReLU activation function in Darknet-53， and the concatenated blocks were used to replace the residual blocks refering to the DenseNet. In the detection head， an Anchor Transformation Net （ATN） was proposed to transforms the initial horizontal anchors into rotated ones. An Anchor Aligned Convolution （AAC） was proposed to adjust the sampling position of convolution operation under the guidance of the rotated anchors. The extracted aligned features were then used to predict the final rotated bounding box and the category of the objects. Experimental results show that the proposed backbone improved the detection accuracy by 1.38%. The proposed AAC feature improved the accuracy by 4.38%， 4.24% and 3.79% respectively compared with the stantard convolution， the deformable convolution and the guided anchoring deformable convolution. Compared with several recent rotated object detection networks， the proposed method can do the detection at a framerate of 21.2 fps while achieving a competitive accuracy as the two-stage detector.

Key words: computer application, UAV aerial image, rotated object detection, deep learning, feature alignment