Research on UAV Perception Systems for Dynamic Indoor Environments: PointNet++ Optimization with Hierarchical Contextual Attention

This research addresses the challenges of dynamic perception systems for UAVs in indoor environments by proposing an optimization method for PointNet++ that integrates hierarchical contextual attention mechanisms. To address the issue of imprecise boundary feature extraction in traditional LiDAR point cloud processing for dynamic object detection and tracking, we designed a Hierarchical Contextual Attention Feature Propagation (HCAFP) module. This module establishes weighted connections between high-level semantic features and low-level geometric features, significantly enhancing the model's perception capabilities for object part boundaries. Experimental results show that the enhanced model achieved an instance average IoU of 86.7% on the ShapeNetCore dataset, an improvement of 1.3 percentage points over the original PointNet++ MSG architecture. Additionally, we proposed a series of model optimization strategies, including model compression and quantization, lightweight model structure design, and hierarchical sparse attention mechanisms, enabling the improved model to be efficiently deployed on UAV platforms with limited computational resources while maintaining competitive performance (85.3% instance average IoU). This research provides a new technical approach for UAV autonomous navigation and dynamic object perception in indoor environments, with significant implications for UAV applications in complex dynamic indoor settings.