Pose Estimation for Weakly Textured Aircraft with Improved CDPN

Obtaining 6d attitude information of a target object from a single RGB image can provide accurate environment awareness and localization capabilities for an aircraft, thereby improving mission execution efficiency and success, while reducing risk and increasing safety. Aiming at the problem that the existing pose estimation algorithms have low accuracy in the small error range in the pose estimation task of weakly textured aircraft, a learning framework based on CDPN network and Convolutional Block Attention Module (CBAM) is proposed. The accuracy of the position estimation of weakly textured aircraft images is effectively improved by increasing the set of predefined keypoints of the wing and tail in order to strengthen the establishment of the correspondence. Experiments have been carried out under the weakly textured aircraft simulation dataset and Linemod dataset, and the results show that the method proposed in this paper has high accuracy within a small error range for weakly textured aircraft images.