Abstract:To address the low pose-estimation accuracy of box-shaped objects in occluded and general environments, a method is proposed that integrates 2D keypoint detection based on RTMPose-BRNM with point-cloud depth information. First, standard convolutions are replaced with RFAConv to improve the localization accuracy of occluded 2D keypoints. Subsequently, the NATTEN module is employed to enhance the model’s capability in extracting edge contour points of box-shaped objects. Furthermore, a mixed-perception convolution (MPC) structure is designed to increase the model’s adaptability to objects of varying sizes. Experimental results show that the proposed RTMPose-BRNM keypoint detection algorithm achieves a mean pixel distance error (MPDE) of 0.98, which is 1.19 lower than that of the original RTMPose model. The improved framework yields translation and rotation errors of approximately 1.32% and 0.96°, respectively.