Researchers develop interpretable AI for UAV target tracking
New fuzzy inference framework achieves 99.7% accuracy in real-time UAV guidance with just 27 rules.
A team of researchers from institutions including North Carolina A&T State University and Amirkabir University of Technology has developed an interpretable fuzzy-inference framework to improve UAV target tracking using bounding-box geometry. The system bypasses the need for complex geometric modeling or large datasets by leveraging low-dimensional features from YOLO detection boxes—target centroid location, area, and aspect ratio.
The framework combines a Mamdani fuzzy baseline with a first-order Takagi-Sugeno model, using a compact 27-rule structure derived from training-set quantiles. Evaluated on 6,169 labeled samples from a VICON motion-capture environment, the Takagi-Sugeno model achieved a mean absolute error of 0.140° ± 0.003°, with 99.676% accuracy within ±1° and 100% within ±3° and ±5°. The system is computationally lightweight, transparent, and suitable for real-time deployment on resource-constrained UAV platforms.
- Uses YOLO bounding-box features (centroid, area, aspect ratio) to generate yaw commands without explicit geometric modeling
- Achieves 0.14° mean absolute error and 99.676% accuracy within ±1° on 6,169 motion-capture samples
- Framework is interpretable, data-efficient, and computationally lightweight—ideal for real-time UAV guidance
Why It Matters
Enables transparent, low-compute AI for UAVs to track moving targets, improving safety and deployment feasibility in real-world robotics.