AI-Powered Video Threat Detection System

Developed an end-to-end computer vision pipeline that ingests and analyzes video footage for hazardous objects. To optimize for constrained compute environments, the system performs frame reduction and color space compression before inference.

Processed frames are evaluated against a custom-trained YOLOv8 model, triggering automated email/SMS alerts upon positive detection.

• Integrated Zipkin to trace the latency of the video processing pipeline (ingest -> compress -> infer -> alert)
• Trained a domain-specific model to reduce false positives compared to off-the-shelf weights