| Abstract |
The segmentation of large-scale urban high vegetation point clouds, typically referring to vegetation communities over 2 meters tall, is crucial for applications in urban ecological management, environmental monitoring, disaster prevention, climate regulation, and smart city development. However, accurately segmenting such vegetation remains challenging due to the irregular and complex nature of point cloud data. To address these challenges, this paper introduces NAM-PointNet++, an enhanced version of PointNet++ [1], which incorporates a Normalized Attention Mechanism (NAM) [2] and differential pooling [3]. The loss function is optimized using Focal Loss to further improve segmentation. NAM enhances key feature extraction, while differential pooling improves the model's ability to capture fine-grained geometric details. Focal Loss effectively mitigates class imbalance. Experiments on the SensatUrban dataset show that NAMPointNet++ surpasses five mainstream deep learning models in urban high vegetation segmentation. © 2024 IEEE. |
