A Real-Time Semantic Segmentation Method Based on STDC-CT for Recognizing UAV Emergency Landing Zones.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• UAV flight safety is critical, necessitating emergency landing capabilities.
• Real-time identification of safe landing zones is essential for autonomous UAVs facing GPS loss or anomalies.

Purpose of the Study:

• To develop a novel semantic segmentation network for accurate and efficient emergency landing zone recognition in UAV aerial imagery.
• To address challenges posed by complex backgrounds, diverse categories, and small targets in UAV imagery.

Main Methods:

• Creation of the UAV-City dataset with 600 annotated aerial images across 12 categories.
• Proposal of the STDC-CT network featuring detail guidance, small object attention, and multi-scale contextual information branches.
• Fusion of feature branches guided by small object attention for improved segmentation.

Main Results:

• The STDC-CT network achieved a balance between segmentation accuracy and inference speed on UAV-City, Cityscapes, and UAVid datasets.
• Achieved 76.5% mIoU at 122.6 FPS on Cityscapes, 68.4% mIoU on UAVid, and 67.3% mIoU at 196.8 FPS on UAV-City.
• Demonstrated real-time performance with an average inference speed of 58.32 ms/image on Jetson TX2.

Conclusions:

• The STDC-CT network effectively recognizes emergency landing zones for UAVs, improving small object segmentation accuracy.
• The proposed method offers a robust solution for real-time semantic segmentation in demanding UAV applications.