A method for deep feature extraction in 3D point cloud semantic segmentation
Addressing the key issues in 3D point cloud semantic segmentation,such as the loss of geometric information,insufficient extraction of local features,and inadequate utilization of local contextual information,this paper proposes a semantic segmentation model that integrates point cloud surface features with spatial visual information,employing dual pooling strategies and self-attention mechanisms to enhance the model's in-depth understanding of point cloud data.Initially,the model captures local geometric structures and spatial layouts with precision through a Curvature and Spatial Visual Local Feature Extraction module,followed by the efficient integration of local features using a Dual Pooling and Self-Attention Local Feature Aggregation module,which strengthens the model's ability to capture global point cloud information.Experimental evaluation on the public S3DIS dataset demonstrates that the model achieves an average Intersection over Union(IoU)accuracy of 71.35％,showing a significant improvement compared to other models.The model also features optimizations in parameter count and computational efficiency,enabling efficient point cloud processing even in environments with limited computing resources.The experimental results confirm that our model not only improves segmentation accuracy but also balances computational efficiency,offering new possibilities for the practical application and widespread use of 3D point cloud semantic segmentation.
point cloudfeature extractionfeature aggregationattention mechanismsemantic segmentation
NETL
NSTL
万方数据
