研究目的
To develop a deep-learning-based approach for classifying terrestrial laser scanner (TLS) point clouds into semantic categories to automate the generation of 3D models for applications like smart cities.
研究成果
The proposed multi-scale hierarchical deep neural network effectively classifies TLS point clouds with high accuracy, outperforming existing methods. It handles varying densities and captures both local and global structures. Future work should focus on automating parameter adjustment and integrating other spatial data for finer classification.
研究不足
The approach requires manual setting of parameters (N, d, S), which may not generalize well to all datasets. Computational cost is high due to the complexity of deep learning and large point cloud sizes. Performance is lower for classes with few samples (e.g., pole and scanning artefacts). Incorporating additional features like color and surface normal did not significantly improve results and added computational overhead.
1：Experimental Design and Method Selection:
The methodology involves a multi-scale hierarchical deep neural network that directly processes 3D point clouds. It uses 3D convolutional neural networks (CNNs) to extract features, subsamples point clouds at multiple scales using VoxelGrid, and employs a softmax regression classifier for semantic labeling. The design aims to handle varying point densities and irregular distributions in TLS data.
2：Sample Selection and Data Sources:
Two benchmark datasets are used: the Oakland dataset (mobile laser scanner data with 5 classes: facade, ground, wire, vegetation, pole) and the Europe dataset (TLS data with 8 classes: man-made terrain, natural terrain, high vegetation, low vegetation, building, hard scape, scanning artefacts, cars). Both datasets are divided into training and testing sets.
3：List of Experimental Equipment and Materials:
A workstation with Intel Xeon E5-2680 CPU (14-core, 2.40 GHz), 64.0 GB RAM, and NVIDIA Quadro P5000 GPU (16.0 GB) is used for implementation. Software includes TensorFlow and Python.
4：40 GHz), 0 GB RAM, and NVIDIA Quadro P5000 GPU (0 GB) is used for implementation. Software includes TensorFlow and Python.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The process includes constructing a kd-tree for nearest neighbor search, feature transformation using symmetric functions, convolutional and max pool layers for feature extraction, subsampling at multiple scales, and classification with softmax regression. Parameters like number of neighboring points (N), depth of hierarchical features (d), and number of scales (S) are optimized through iterative testing.
5：Data Analysis Methods:
Performance is evaluated using precision, recall, F1 score, overall accuracy, and Intersection over Union (IoU). Cross-entropy loss function is minimized using backpropagation for parameter optimization.
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
