Project
Warehouse Parcel Monitoring System
A warehouse video-analytics pipeline for parcel-condition monitoring, OCR-assisted metadata extraction, movement tracking, and incident review.
Why I Built It
Warehouses generate huge amounts of video but most of it is never reviewed unless something goes wrong. I wanted to build a pipeline that actively monitors parcel condition in real time and makes it easy to review incidents after the fact.
The Problem
Warehouses need practical ways to detect parcel damage, track movement, and review incidents from operational video feeds.
The Solution
The pipeline combines object detection, OCR-assisted metadata extraction, movement tracking, and structured incident review APIs.
Engineering Decisions
Chose YOLOv8 over Transformer-based detectors for inference speed on edge hardware. Used MLflow for experiment tracking across 50+ training runs. DVC for dataset versioning to ensure reproducible evaluations.
Interesting Challenges
Problem
Balancing detection accuracy with inference latency for real-time processing of 30 FPS warehouse footage on modest GPU hardware.
Solution
Optimized the pipeline with frame sampling (every 5th frame) and model quantization (FP16), reducing latency from 120ms to 45ms per frame while maintaining 96%+ mAP50.
Outcome
Real-time processing at 6.6 effective FPS on an RTX 3060, sufficient for operational monitoring without dropping critical events.
Problem
OCR on parcel labels captured from moving cameras at varying angles produced inconsistent text extraction.
Solution
Implemented a multi-frame voting scheme that aggregates OCR results across consecutive frames, discarding outliers before assembling the final text.
Outcome
Label extraction accuracy improved from 74% to 91% on the evaluation set.
Results
The parcel detection model reached 92.7% precision, 95.0% recall, and 97.4% mAP50 in evaluation.
What This Demonstrates
Computer vision pipeline engineering, model evaluation rigor, MLOps practices (MLflow, DVC), and practical deployment-aware optimization for real-time video analytics.
Limitations
Performance depends on camera angle, lighting, label quality, and the diversity of parcel conditions in training data.
Future Improvements
Improve multi-camera tracking, active learning loops, and dashboard workflows for operations teams.