SWUAV-DANet: A Severe-Weather UAV Dataset and Dynamic AlignAir Network for Robust Aerial Vehicle Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Unmanned aerial vehicle (UAV) aerial object detection is crucial for applications like traffic monitoring and emergency response.
• Existing methods struggle with performance degradation in adverse weather (heavy rain, fog, blizzards) and poor lighting due to information loss and feature instability.

Purpose of the Study:

• To address the challenges of vehicle detection in severe weather conditions using UAVs.
• To introduce a novel dataset and a real-time detection network designed for adverse environments.

Main Methods:

• Construction of the Severe-Weather UAV (SWUAV) dataset, featuring 18,195 RGB aerial images across 12 adverse conditions with 236,392 vehicle instances.
• Proposal of the Dynamic AlignAir Network (DANet), incorporating a cross-scale adaptive alignment module for feature enhancement and a dynamic adaptive alignment head (DAAH) for improved detection.
• The cross-scale module utilizes adaptive channel calibration, contrastive self-attention, and geometric/semantic remapping to handle scale variations and noise.

Main Results:

• DANet improved the YOLOv11-s baseline's average precision (AP) from 43.9% to 46.9% and AP50 from 62.6% to 64.8% on the SWUAV dataset.
• The network achieved high efficiency with 8.65 million parameters and a throughput of 323.47 frames per second (FPS).
• DANet demonstrated superior performance compared to existing models like EdgeYOLO-s and RT-DETR.

Conclusions:

• The proposed SWUAV dataset and DANet provide a significant advancement for reliable vehicle detection in challenging weather conditions.
• DANet offers a real-time, efficient, and accurate solution for UAV-based aerial object detection in adverse environments.
• The publicly available dataset and code facilitate further research and development in this domain.