Abstract: Extracting a local feature descriptor of point cloud with high description and strong robustness is a key step in point cloud registration. For the problems of existing learning-based descriptors not having rotation invariance or relying on hand-crafted features that are sensitive to noise, a 3D local feature descriptor generation network based on dynamic graph convolution and PointNet is proposed to extract local feature descriptors with rotation invariance and strong generalization. Firstly, the local patch aligned with the local reference frame is used as the input to the network, and the local geometric features and point features are extracted by the dynamic graph convolution model and the PointNet model, respectively, to solve the problem that a single PointNet model is unable to learn the relationship between points in the input point set. Then, to further improve the learning ability of the network, a dual attention mechanism layer, containing a point self-attention module and a local spatial-attention module, is proposed to better integrate the features extracted by the two models to obtain the final descriptor features. Extensive experiments on the indoor dataset 3DMatch and the outdoor datasets ETH and KITTI show that the proposed network achieves a feature matching recall of 98.2% on 3DMatch, and a feature matching recall and success rate of 98.7% and 99.82% on ETH and KITTI, respectively, verifying the effectiveness of the proposed method.