Point cloud is an important three-dimensional data representation applied in fields such as driverless driving and virtual reality. Due to the high resolution of point clouds, data transmission requires a large amount of network bandwidth and storage resources, which seriously hinders further promotion. This aims at the problem that most current point cloud geometric compression algorithms based on deep learning only consider the local feature information or global feature information of the point cloud separately. To this end, a point cloud geometric compression network based on dual-branch feature fusion is proposed. First, to capture local features with rotation invariance, the point cloud resnet module is introduced to extract local features with low computational complexity. Second, To capture the high-dimensional information in the compression process, a Transformer for point clouds is designed. At the same time, to further effectively utilize features, a point cloud attentional feature fusion block is designed, which contains the point cloud channel attention blocks for local and global features effective fusion extraction. Finally, To compensate for the feature loss in the sampling process and reduce the dynamic memory footprint of model training, the autoencoder adopts a multi-scale progressive structure. Experimental results demonstrate that on the 8iVFB, Owlii, and MVUB datasets, which average Bjontegaard Delta Rate (BD-Rate) gains of 89.11% and 88.17% compared to Geometry-based Point Cloud Compression (G-PCC).