Abstract: The large amount of point cloud data, unbalanced distribution and uneven density of urban road scenes make it difficult for the current point cloud classification methods to meet the requirements of high-precision classification. To deal with the problem of insufficient local feature extraction by PointNet++ networks, a local feature aggregation module is designed based on the attention mechanism, which adequately captures local information by dynamically merging neighboring point features. Considering that the existing classification models cannot take into account contextual information, which leads to limited classification performance in complex scenes, a dual-attention module and a context-aware module are constructed to extract contextual information from several dimensions to further enhance the feature representation capability. The experimental results show that the new method has higher accuracy and stronger generalization performance (overall accuracy reaches 98.70% and 96.84% in Oakland and Paris publicly available datasets) under large point cloud datasets, and is more suitable for large-scale point cloud classification.

Key words: point cloud classification, PointNet++, local feature, attention mechanism, contextual information