Overview
This system, developed for the TEKNOFEST 2025 Combat UAV Competition, enables autonomous aerial vehicles to detect and track other UAVs in the air in real time, and then produce autonomous guidance decisions from this information.
System Components
Detection Layer
- Fine-tuned model on YOLOv8
- Custom dataset: A manually annotated training set of UAV images
- Real-time inference (~30 FPS target)
Tracking Layer
- DeepSORT: Object identity tracking with visual features
- Kalman Filter: Future position prediction with motion model
- Centroid Tracking: Backup mechanism for fast, simple tracking
- Multi-target management (multiple UAVs simultaneously)
Decision Layer
- Creating a density map from detection data
- Marking regions with high detection density as danger zones
- Autonomous guidance decision: "Head to this region" / "Track this target"
Technical Challenges
- Low latency requirement: Image → detection → decision chain must stay sub-100ms
- Noisy environment: Filtering false positives like clouds and birds in outdoor footage
- Target loss: Continuing tracking with Kalman prediction when the UAV temporarily leaves the camera frame
TEKNOFEST Result
We applied as participants to the TEKNOFEST 2025 Combat UAV Competition; the system worked successfully in competition-oriented test environments.