Execution Checklist

Phase 0: Current Baseline

• Review the legacy repository
• Create clean radar restart plan
• Create multimodal fusion plan
• Build Project v1
• Add scenario runner
• Add tutorial notebook
• Add radar TUI
• Add Obsidian project documentation
• Align the audio runtime with drone_thesis_audio_training.ipynb

Phase 1: Validate Project v1

• Install Project v1/requirements.txt
• Run predefined radar scenarios locally
• Run the radar TUI locally
• Run audio.video_test locally with a dependency-complete environment
• Record observed failures and mismatches
• Tune radar-only baseline until stable

Phase 2: Improve Radar Reliability

• Add M/N temporal confirmation
• Add simple track persistence
• Compare CA-CFAR with stronger alternatives if needed
• Quantify radar-only false alarms
• Create clutter-stressed evaluation scenario

Phase 3: Vision Module

• Freeze detector choice
• Freeze tracker choice
• Define video input interface
• Define per-frame/per-track output schema
• Train or integrate the selected detector
• Measure vision-only precision/recall

Recommended vision path

• Start with YOLO11
• Add ByteTrack or BoT-SORT style tracking
• Export time-window confidence scores

Phase 4: Audio Module

• Freeze the active audio artifact as drone_sound_model.h5
• Match runtime preprocessing to the training notebook
• Define time-window confidence output
• Record the final validation metrics for the retained CNN
• Measure audio-only precision/recall
• Calibrate threshold and M/N confirmation settings on labeled clips

Recommended audio path

• Use Notebooks/drone_thesis_audio_training.ipynb as the retained training source
• Use src/audio/drone_sound_model.h5 as the active runtime artifact
• Document the final dataset split and label policy
• Decide final production settings from validation results

Phase 5: Fusion Layer

• Define common UnifiedDetectionEvent
• Define timestamp alignment strategy
• Define score normalization strategy
• Implement weighted late fusion
• Implement M/N fused alert logic
• Log per-modality contribution in fused alerts

Phase 6: Evaluation

• Prepare negative cases
• Prepare positive cases
• Run radar-only ablation
• Run radar + vision ablation
• Run radar + audio ablation
• Run full multimodal ablation
• Compare false alarms per minute
• Compare fused precision/recall/F1

Phase 7: Final Presentation

• Export final figures
• Create final notebook/report version
• Create final system architecture diagram
• Create defense summary slide content
• Write limitations section clearly

Linked Notes

• Dashboard
• Current Status
• Multimodal Fusion Roadmap
• Project V1 Codebase