Spatio-Temporal Feature Fusion for Anti-UAV Detection: Integrating Inter-Frame Dynamics and Appearance.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Detecting small, slow-moving unmanned aerial vehicles (UAVs) in cluttered environments poses significant challenges.
• Existing methods often struggle with low visibility, complex backgrounds, and subtle motion detection.

Purpose of the Study:

• To develop an advanced detection system for low-slow-small UAV targets in complex scenarios.
• To improve the precision, recall, and mean average precision (mAP) of UAV detection.
• To create a practical and efficient system deployable on embedded hardware.

Main Methods:

• An improved YOLOv11 static detector incorporating SPD Conv, BiFPN, and a high-resolution detection header.
• A dynamic target-detection algorithm to capture subtle movement features.
• An integrated strategy fusing static and dynamic detection judgments.

Main Results:

• The proposed MSM-YOLO method achieved Precision of 94%, Recall of 92%, and mAP50 of 86.3%.
• Significant improvements over the baseline YOLOv11 detector, with increases of 12.1% in Precision, 29.5% in Recall, and 29.6% in mAP50.
• Ablation studies confirmed the effectiveness of individual modules.
• Optimized deployment on an RK3588 embedded system achieved 100 frames per second (fps).

Conclusions:

• The novel spatio-temporal information fusion method effectively enhances the detection of small UAVs in complex backgrounds.
• MSM-YOLO demonstrates superior performance and practicality for real-world air-to-air UAV detection applications.
• The system's efficiency and accuracy make it suitable for deployment on resource-constrained embedded platforms.