| Popis: |
The semantic segmentation of laser point clouds is critical for many applications of aerial point clouds. However, most of the existing deep learning networks do not make full use of point cloud data information. PointNet++ was chosen as the baseline network, and a deep-residual enhanced encoding method of multi-feature information is proposed in this work. Firstly, a more efficient network structure to enhance geometric information encoding is constructed, called the GEO–PointNet layer. Then, a novel structure for feature aggregation, named SEP–PointNet, is introduced to encode the auxiliary and geometric features of points separately. Additionally, the segmentation network is deepened in the way of residual structures, which can effectively restrain network degradation. Meanwhile, ‘Dropout’ operations are applied to the fully connected layer to cope with the problem that the model is prone to overfitting due to many network parameters. Finally, a novel segmentation network, named SGDD–PointNet++, is built, and its effectiveness was evaluated by using four airborne benchmark datasets. The experimental results performed on the DALES dataset indicate that the overall accuracy and average intersection-over-union (mIoU) value of the modified PointNet++ networks are better than the original baseline and the other two state-of-the-art segmentation methods. The overall accuracy of the improved SGDD–PointNet++ network reached 87.88%. For the category IoU, it also outperforms other networks, and it has a maximum accuracy increment of 11.43%. Meanwhile, in terms of the generalization capabilities of the trained models, the proposed network can provide better discrimination effects for three public aerial datasets than other methods. |
