Abstract: The monocular 6D pose estimation method based on two-stage coordinate decoupling has the characteristics of stability, high efficiency and fast training speed, but there is still room for improvement in accuracy. A monocular 6D pose estimation algorithm using Gaussian heatmap coordinate regression and fusion attention is proposed. It introduces the fusion attention module into resnet34 backbone network, so that the network can better learn the surface features and spatial information of objects. Based on the differential space coordinate transformation, the translation calculation network is improved to predict the coordinate translation more accurately. The algorithm uses a clustering method based on density hierarchy and establishes a hash point cloud index, constrain the predicted 3D point cloud, and effectively reduce the outliers of 3D sampling points. In the training phase, the algorithm uses the synthetic rendered image to expand the linemod data set and provide rich data for network training. The experimental results show that the ADD(-S) index and 2D projection error index of the method reach 93.27% and 98.81% respectively. Compared with the benchmark algorithm CDPN, it is improved by 3.41% and 0.79% respectively. Compared with the relatively novel algorithms such as PVNet and DPOD, it shows comprehensive advantages.