Real-Time Topology-Aware Branch Segmentation for UAV Perception in Natural Environments.

Autonomous UAVs operating in forest environments require accurate branch perception to enable reliable environmental understanding for downstream tasks such as navigation, environmental inspection, and obstacle avoidance. However, existing semantic segmentation methods primarily focus on pixel-wise classification and often fail to preserve the structural topology of branches, making it difficult to distinguish reliable branch segments from complex regions such as junctions and overlapping structures. In addition, achieving robust real-time performance on resource-constrained onboard platforms remains challenging due to the slender and irregular characteristics of natural branches. To address these challenges, this paper proposes a topology-aware branch perception framework that integrates real-time efficient semantic segmentation with structural topology analysis. For branch segmentation, a Strip-Swift Pyramid Pooling Module (SSPPM) is introduced to enhance elongated structure representation through progressive pooling and strip-based directional context aggregation. A reparameterized Golden Cudgel Block (GCBlock) and a Boundary Optimization Module (BOM) are further incorporated to improve deployment efficiency and boundary quality. Building upon the predicted segmentation masks, a topology-aware structural analysis module is developed to identify and remove branch junctions and multi-branch regions through skeleton-based connectivity analysis, preserving only structurally stable branch segments for subsequent perception and analysis. This strategy improves the robustness and consistency of branch representation in cluttered natural environments. Experimental results demonstrate that the proposed method achieves 89.94% mIoU on the Drone-Branch dataset. When deployed on the NVIDIA Jetson Orin Nano with TensorRT acceleration, the proposed segmentation network achieves a stable inference latency of 13.7 ms. The topology-aware structural analysis is subsequently applied as a post-processing stage to improve the reliability of extracted branch structures. These results demonstrate the effectiveness of the proposed framework for branch perception on resource-constrained UAV platforms.