A 6D Object Pose Estimation Method combining Self-attention Mechanism

Aiming at the accuracy issue of the 6D pose estimation algorithm, a YOLO-6D object pose estimation method which introduces the expectation maximization self-attention mechanism is proposed. Firstly, use CSPDarkNet53 network to replace the original DarkNet19 network, increase network depth, and improve the network's feature extraction ability. Secondly, the expectation maximization self- attention module (EMAU) is combined with residual structures to propose the ResEMA structure, which is incorporated into the backbone network to extract more fine-grained representations. Finally, experiments are conducted on the publicly available dataset LineMOD, and the results show that the improved algorithm significantly improves accuracy, with a 6.78% increase in 2D projection error accuracy, a 9.14% increase in ADD accuracy, a 13.44% increase in 5cm,5 degrees accuracy.