Aerial small object detection via dynamic convolution and hierarchical attention fusion for UAV imagery

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Small object detection in Unmanned Aerial Vehicle (UAV) aerial imagery faces challenges due to small object size, limited features, and complex backgrounds.
• Existing methods struggle to balance accuracy and computational efficiency for real-time aerial surveillance.

Purpose of the Study:

• To develop an advanced aerial small object detection algorithm, AeroVision-DET, for improved performance on UAV platforms.
• To address limitations in feature representation and fusion for detecting small objects in complex aerial scenes.

Main Methods:

• Introduced the Adaptive Receptive Field Network (ARFNet) backbone with Multi-Scale Adaptive Feature Modules (MSAFM) for adaptive receptive field adjustment.
• Designed the Semantic-Spatial Fusion Module (SSFM) utilizing multi-scale contextual attention for robust feature fusion.
• Proposed the Efficient Feature Encoding Layer (EFEL) with polarized linear attention and frequency-modulated networks for efficient long-range dependency modeling.

Main Results:

• AeroVision-DET achieved 24.2% AP on the VisDrone2019 dataset, outperforming the RT-DETR-r18 baseline by 3.4% and improving small-object AP by 3.5%.
• The algorithm reduced parameters by 27.0% with similar computational complexity, achieving 69.5 FPS real-time inference.
• Demonstrated competitive detection accuracy with significantly fewer parameters (14.57 M) and lower computational cost (56.9 GFLOPs).

Conclusions:

• AeroVision-DET offers a highly efficient and accurate solution for small object detection in UAV aerial imagery.
• The algorithm is suitable for deployment on resource-constrained edge devices, enabling real-time aerial surveillance applications.
• The proposed methods effectively enhance feature representation and fusion for challenging small object detection tasks.