Real-time Firearms Detection

An HEC-funded research project required real-time firearm detection in video streams. Both accuracy and inference speed mattered for live camera feeds.

Existing detectors were either too slow for real-time use or not accurate enough under varying angles and occlusion.

I implemented real-time firearms detection using YOLOv3, optimized for live webcam footage with model and inference optimizations for usable frame rates.

Python and PyTorch power the pipeline with Docker and GPU tuning for reproducible, low-latency inference.

Fully dockerized GPU deployment ensures reproducible inference across environments. Documented pipeline for future research and deployment.

Achieved over 90% accuracy with optimized inference suitable for real-time use.

Reproducible Docker setup and documented pipeline for future research and deployment.