Intelligent 3D Reconstruction Algorithms for Low-Altitude Surveying Based on Deep Point Cloud Learning
Artificial intelligence-driven deep point cloud learning technology offers novel solutions for low-altitude surveying and mapping 3D reconstruction. Addressing the accuracy limitations of traditional reconstruction algorithms under conditions of sparse point clouds, pose drift, and noise interference, this study constructs a deep point cloud learning-based intelligent 3D reconstruction model for low-altitude surveying. The model employs a multi-scale feature extraction network as its foundation, integrating spatial alignment and density compensation mechanisms to achieve robust feature fusion. Building upon this, an attention-guided generative adversarial network (GAN) architecture is introduced to enhance local detail reconstruction and global topological consistency. The feature aggregation module employs progressive neighborhood sampling and nonlinear mapping to establish hierarchical geometric and semantic correlations across network layers. The generative network’s multi-objective loss function integrates Chamfer Distance, Earth Mover’s Distance, and adversarial loss, achieving dual optimization of reconstruction quality and computational efficiency. Experiments validated the model on self-collected low-altitude LiDAR data and the Semantic3D dataset. Compared to mainstream algorithms like PointNet++, DGCNN, and PCN, our proposed model reduces Chamfer Distance by approximately 28.9%, improves F1-score by about 3.6%, and shortens average inference time to 43.8 ms/frame. Results demonstrate that this model achieves high-precision, low-noise, structurally complete 3D reconstruction in complex low-altitude environments, exhibiting strong generalization and engineering adaptability.