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

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous Unmanned Aerial Vehicles (UAVs) require precise branch perception for forest environment understanding.
• Existing semantic segmentation methods struggle with branch topology preservation and real-time performance on resource-constrained platforms.
• Challenges include distinguishing complex branch structures like junctions and overlapping regions.

Purpose of the Study:

• To develop a topology-aware branch perception framework for autonomous UAVs.
• To integrate efficient semantic segmentation with structural topology analysis for improved branch representation.
• To enable robust real-time branch perception on resource-constrained UAV platforms.

Main Methods:

• Introduced a Strip-Swift Pyramid Pooling Module (SSPPM) for enhanced elongated structure representation.
• Incorporated a reparameterized Golden Cudgel Block (GCBlock) and Boundary Optimization Module (BOM) for efficiency and boundary quality.
• Developed a topology-aware structural analysis module using skeleton-based connectivity analysis for junction and multi-branch region removal.

Main Results:

• Achieved 89.94% mean Intersection over Union (mIoU) on the Drone-Branch dataset.
• Demonstrated a stable inference latency of 13.7 ms on an NVIDIA Jetson Orin Nano with TensorRT acceleration.
• Post-processing structural analysis improved the reliability of extracted branch structures.

Conclusions:

• The proposed framework effectively addresses challenges in real-time branch perception for UAVs in forest environments.
• The integration of semantic segmentation and topology analysis enhances robustness and consistency of branch representation.
• The method is suitable for deployment on resource-constrained UAV platforms for critical downstream tasks.