Point cloud semantic segmentation presents unique challenges due to the irregular, unordered nature of 3D spatial data. To address this, we developed the Parallel Hybrid Attention Mechanism (PHAM) integrated with PointNet++, creating PHAM-PointNet++ for enhanced feature extraction. This architecture processes channel and spatial information in parallel through dual attention pathways:
The channel attention mechanism refines feature discrimination using max-pooling and average-pooling operations combined with multilayer perceptrons (MLPs):
$$M_c = \sigma(W_2(ReLU(W_1(F_{max}))) + W_2(ReLU(W_1(F_{avg})))) \otimes F$$
where $F_{max}$ and $F_{avg}$ represent max-pooled and average-pooled features, $W$ denotes MLP weights, and $\otimes$ indicates element-wise multiplication.
Spatial attention enhances geometric relationship understanding:
$$M_s = \sigma(f(F_{max}, F_{avg})) \otimes F$$
where $f(\cdot)$ transforms pooled features into spatial weights.
The outputs are fused with original features through residual connections, preserving critical input characteristics while amplifying discriminative elements. This architecture processes both feature channels and spatial relationships simultaneously, significantly improving recognition of structural patterns essential for low altitude UAV navigation in complex environments.
For low altitude drone path planning, segmented point clouds are converted into navigable 3D occupancy grids. Point coordinates are normalized using the Rescale function for consistent spatial mapping:
$$Y = (X – \min(X)) \frac{(N_{new\_max} – N_{new\_min})}{(\max(X) – \min(X))} + N_{new\_min}$$
where $X$ represents raw coordinates and $Y$ denotes normalized values. Each grid cell stores position and semantic labels (Table 1), with obstacle identification triggering binary occupancy flags (0 = free, 1 = occupied).
| Semantic Label | Category | Color (RGB) |
|---|---|---|
| 200800 | Ground | [0.68, 0.85, 0.90] |
| 200900 | Others | [0.77, 0.93, 0.49] |
| 200101 | Building | [0.53, 0.81, 0.98] |
| 200301 | Tree | [0.75, 0.75, 0.00] |
| 200400 | Veg | [0.31, 0.58, 0.80] |
We employed the IGWO-WF (Improved Grey Wolf Optimizer with Weighting Functions) algorithm for 3D trajectory generation. This bio-inspired approach simulates wolf pack hunting behavior with position vectors dynamically updated using weighting functions to minimize path length while maintaining safe clearance from obstacles identified in the occupancy grid. The optimization objectives for low altitude UAV navigation include:
$$J_{total} = \alpha \cdot J_{length} + \beta \cdot J_{clearance} + \gamma \cdot J_{smoothness}$$
where $\alpha$, $\beta$, and $\gamma$ balance path length, obstacle distance, and flight smoothness.
Experimental validation on S3DIS and WHU-Urban-3D datasets demonstrates PHAM-PointNet++’s superiority. The model achieved significant improvements in segmentation metrics critical for low altitude drone perception systems (Table 2).
| Method | mIoU (%) | Accuracy (%) |
|---|---|---|
| PointNet++ | 73.06 | 92.51 |
| CBAM-PointNet++ | 73.09 | 92.43 |
| PHAM-PointNet++ | 75.01 | 92.69 |
Specific class improvements highlight the model’s enhanced capability to identify navigation-critical features: Water (2.80% IoU increase), Vehicle (4.66% IoU increase), and non-Vehicle (8.25% IoU increase) classes showed the most significant gains. These improvements directly enhance obstacle identification accuracy for low altitude UAV operations in urban environments.
For trajectory planning validation, we generated a 100m×100m×100m occupancy grid from segmented point clouds. The IGWO-WF algorithm produced collision-free paths averaging 123.86m length with smooth altitude transitions. Figure 1 demonstrates successful obstacle avoidance across complex structures including buildings, vegetation, and elevated infrastructure – critical capabilities for low altitude drone operations in constrained airspace.
The Nesterov Accelerated Gradient (NAG) optimizer enhanced training efficiency through momentum-based parameter updates:
$$v_t = \mu v_{t-1} – \eta \nabla f(\theta_{t-1} + \mu v_{t-1})$$
$$\theta_t = \theta_{t-1} + v_t$$
where $\mu$ controls momentum and $\eta$ represents the learning rate. Comparative experiments showed NAG outperformed Adam and SGD optimizers by 11.58% mIoU and 3.06% accuracy respectively when training PHAM-PointNet++, substantially accelerating convergence for low altitude UAV perception models.
This integrated approach demonstrates practical viability for autonomous low altitude UAV inspection systems. The PHAM-PointNet++ architecture provides robust environmental understanding while the IGWO-WF algorithm enables efficient 3D path planning in obstacle-dense environments. Future work will focus on real-time implementation and multi-sensor fusion to enhance operational reliability for complex low altitude drone missions.