Hybrid AI Drone Detection System -- 2
Budget: $30 – $250 USD
I need a working proof-of-concept that spots drones in real time by blending vision and RF cues. On the vision side you will train and fine-tune YOLOv8 with Anti-UAV and VisDrone footage. In parallel, spectrograms coming from DeepSig RadioML together with my own SDR captures (HackRF / RTL-SDR) should feed a CNN that flags drone-class emissions.
Once both streams run reliably, I want them merged—either with a straightforward rule set or a small neural fusion layer; I’m happy to discuss which choice achieves the best balance of speed and robustness. The finished model must be benchmarked on four fronts that matter equally to me: Accuracy, Precision, Recall, and False-Alarm Rate.
Because the project will eventually guard a sensitive site, I’m aiming for a solution that can move from laptop to field hardware without major rework. Python, PyTorch, and standard CV/RF libraries are expected; please keep your code clean, well-commented, and reproducible.
Deliverables:
• Trained YOLOv8 weights and training notebook
• Trained RF-CNN model and training notebook
• Fusion module with documented logic or architecture
• End-to-end inference script that ingests live or recorded video plus I/Q data and outputs detection events
• Performance report covering the four metrics above, including confusion matrices and a short discussion of failure cases
Timeline is tight—I’d like initial results as soon as possible, so only reply if you can start right away and iterate quickly.
Once both streams run reliably, I want them merged—either with a straightforward rule set or a small neural fusion layer; I’m happy to discuss which choice achieves the best balance of speed and robustness. The finished model must be benchmarked on four fronts that matter equally to me: Accuracy, Precision, Recall, and False-Alarm Rate.
Because the project will eventually guard a sensitive site, I’m aiming for a solution that can move from laptop to field hardware without major rework. Python, PyTorch, and standard CV/RF libraries are expected; please keep your code clean, well-commented, and reproducible.
Deliverables:
• Trained YOLOv8 weights and training notebook
• Trained RF-CNN model and training notebook
• Fusion module with documented logic or architecture
• End-to-end inference script that ingests live or recorded video plus I/Q data and outputs detection events
• Performance report covering the four metrics above, including confusion matrices and a short discussion of failure cases
Timeline is tight—I’d like initial results as soon as possible, so only reply if you can start right away and iterate quickly.