It is known that PointNet++ cannot deeply explore the semantic features of Lidar point clouds during feature extraction and loses features due to the use of maximum pooling during feature aggregation. These problems result in a decrease in point cloud segmentation accuracy. To address these problems, this study proposes a point cloud segmentation model based on feature deviation values and attention mechanisms by improving the feature extraction and feature aggregation modules of PointNet++. First, different local neighborhoods are obtained using spherical sampling, and then neighborhood points are selected using the K-nearest neighbor (KNN) method to calculate the feature deviation values of different neighborhoods. This enhances the model's recognition ability for different local neighborhoods and obtains deep semantic information of the point cloud. Second, an attention-based feature aggregation module is used to replace the maximum pooling module in PointNet++ to learn the weights of different features during feature aggregation. This improves the model's ability to filter information from different structures and enhances the segmentation performance of the model. To further optimize the model architecture, a residual module is added to the fully connected layer to avoid parameter redundancy and improve the model performance through weight sharing. Experimental results are validated on the Vaihingen and S3DIS datasets provided by ISPRS and Stanford University, respectively, and are compared with the results of the experiments and mainstream models provided by ISPRS. The overall accuracy (OA) and average F1 score reach 86.69% and 73.97%, respectively, which are 5.49 percentage points and 8.30 percentage points higher than those of PointNet++, respectively. The experimental results on the S3DIS dataset are compared with those of PointNet++, RandLA-Net, and ConvPoint, and show a clear improvement over those of PointNet++. The experiments show that the improved model can fully extract the semantic features of point clouds and effectively improve the segmentation accuracy of the model as compared with the segmentation results of PointNet++.