研究目的
Development of a deep convolutional neural network (CNN) architecture for automatic classification of outdoor mobile LiDAR data, addressing challenges like noise, clutter, and large information quantum.
研究成果
The proposed CNN architecture successfully classifies outdoor MLS data with high accuracy, overcoming challenges like noise and clutter. The LUT based approach enhances classification accuracy by preserving geometry. Future work includes incorporating more spatial features and improving the architecture for higher accuracy on larger datasets.
研究不足
Hardware constraints limited the number of feature maps and voxels that could be used, potentially affecting the classification accuracy. The study also faced challenges in extracting a large number of samples from MLS for other classes.
1：Experimental Design and Method Selection:
The study employs a deep CNN architecture with a look up table (LUT) based approach for rescaling MLS data while preserving geometry. The architecture includes five convolutional layers and five fully connected layers.
2：Sample Selection and Data Sources:
MLS datasets from Streetmapper and Lynx sensors were manually segmented to extract tree and non-tree samples. Additional samples from the Modelnet dataset were used for testing.
3：List of Experimental Equipment and Materials:
Intel Xeon 3.5 GHz processor with 12 cores and NVIDIA Quadro K4000 GPU were used for training the CNN.
4：5 GHz processor with 12 cores and NVIDIA Quadro K4000 GPU were used for training the CNN.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The input MLS point cloud is rescaled using an LUT approach, voxelized into 20x20x20 voxels, and classified using the proposed CNN architecture.
5：Data Analysis Methods:
Classification accuracy is evaluated in terms of total accuracy and kappa values.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
