DYNA JOURNAL ENGINEERING

Key words:

Estimación de pose 6D, YOLOv5-6D, PANet, CBAM, Transformador, PnP, 6D Pose Estimation, YOLOv5-6D, PANet, CBAM, Transformer, PnP

Article type:

ARTICULO DE INVESTIGACION / RESEARCH ARTICLE

Section:

RESEARCH ARTICLES

Six degree-of-freedom (6D) pose estimation is essential for robotic grasping, autonomous driving, and augmented reality, aiming to predict the three-dimensional (3D) translation and rotation of objects. Traditional template- and feature-based methods lack robustness under occlusion and illumination changes. Although YOLO-based approaches (e.g., YOLO6D, YOLOv5-6D) improve efficiency, they still suffer from weak feature modeling and limited multi-scale fusion. To address this, we enhance YOLOv5-6D by introducing path aggregation network (PANet) for stronger
feature fusion, convolutional block attention module (CBAM) for attention-guided feature extraction, and a transformer module
for global context modeling. The final 6D pose is computed using the perspective-n-point (PnP) algorithm. Experiments on the LINEMOD
dataset show that our method improves ADD accuracy from 71.63% to 74.22% while maintaining real-time performance, demonstrating improved accuracy and robustness for practical 6D pose estimation. .