1.8 KiB
Project Proposal: Attention-Enhanced 3D Point Cloud Segmentation for Railway Foreign Object Detection
Research Question and Hypothesis
Research Question: Can attention-based 3D point cloud segmentation architectures significantly improve the detection of rare foreign objects on railway tracks compared to standard RandLA-Net?
Hypothesis: A hybrid architecture combining Attention mechanisms with RandLA-Net's efficient sampling will achieve at least 65% IoU for foreign object detection, while maintaining computational efficiency suitable for real-time monitoring.
Literature Review
Recent advancements in transformer-based architectures like Point Transformer V3 [Wu et al., 2024] have shown superior performance in capturing fine-grained features. Meanwhile, approaches like Sphere Transformer [Lai et al., 2023] specifically address LiDAR's unique distribution patterns. The class imbalance problem has been partially addressed by RAPiD-Seg [Li et al., 2024] through range-aware feature embeddings.
Methodology
We will develop a hybrid architecture integrating:
- RandLA-Net's efficient random sampling for computational feasibility
- Point Transformer's vector attention mechanisms for capturing fine details
- Class-balanced loss functions to address the extreme imbalance (0.001% foreign objects)
- Data augmentation techniques including selective jittering [Park et al., 2024]
Evaluation will use precision, recall, and IoU metrics, with special emphasis on foreign object performance. Computational efficiency will be benchmarked against RandLA-Net baseline.
Planning
- Weeks 1-2: Architecture design and implementation
- Weeks 3-4: Training and hyperparameter optimization
- Weeks 5-6: Evaluation and comparison with baseline
- Weeks 7-8: Documentation and reporting