Automated Determination of Object Pose in a 3D Scene Under Limited Observability Based on Keypoint Analysis

Abstract

The study addresses automated determination of object pose in a three-dimensional scene under limited observability based on keypoint analysis, integrating high-speed image detection, geometric pose recovery, and scientific visualization of the results. The proposed pipeline includes detection of 2D keypoints with visibility assessment, a robust Perspective n Point procedure for estimating extrinsic parameters, and interactive 3D visualization aligned with the camera parameters. The computational scheme relies on OKS/mAP and PCK to assess localization accuracy, IoU and reprojection error to evaluate reconstruction quality, and FPS to analyze performance. The experimental section demonstrates mAP50=0.87 and PCK@5px=0.92 under partial occlusions, enabling stable pose recovery when 6–8 points are correctly observed, while the final 3D reconstruction accuracy reaches IoU=0.94 with a mean reprojection error of 2.3 pixels. The pipeline achieves 146 fps at the detection stage and about 40 fps for the full “image → pose → 3D visualization” cycle, meeting real-time application requirements, including robotics and AR scenarios. The discussion highlights the method’s advantages under limited observability and its independence from depth data, as well as limitations under extreme occlusions and lighting; future directions include self adaptation to new object classes, extension to dynamic scenes, and the use of event sensors to improve robustness.