AAB-FusionNet: A real-time object detection model for UAV edge computing platforms.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned aerial vehicles (UAVs) face computational limitations for real-time object detection.
• Cluttered backgrounds and small/occluded targets pose significant challenges for current UAV systems.

Purpose of the Study:

• To introduce AAB-FusionNet, a novel real-time object detection model optimized for UAV edge computing.
• To enhance detection accuracy and efficiency for small or partially occluded objects in aerial imagery.

Main Methods:

• Developed the Adaptive Attention Block (AAB) with Adaptive Saliency-based Attention (ASA) for focused feature extraction.
• Implemented a Multi-layer Feature Fusion Network integrating Attentive Inverted Bottleneck Aggregation (AIBA), DySample, and Dual-Attention Noise Mitigation (DNM).
• Designed for real-time inference on resource-constrained UAV edge platforms.

Main Results:

• AAB-FusionNet demonstrated robust detection performance, particularly for small and occluded objects.
• Achieved real-time inference speeds on low-power UAV hardware.
• Effectively balanced accuracy, computational efficiency, and adaptability in complex aerial scenarios.

Conclusions:

• AAB-FusionNet provides a highly effective solution for real-time object detection on UAVs.
• The model's adaptive attention and feature fusion mechanisms improve robustness in challenging conditions.
• It is well-suited for applications requiring fast, reliable object identification in aerial surveillance and monitoring.