Case study · 03 / 2023 · Computer Vision
Real-time object detection and tracking for fertilizer classification on a moving production line — built and tuned for Ca Mau Fertilizer Corporation factory conditions.

- Role
- Computer Vision
- Year
- 2023
- Outcome
- ~92% Accuracy · Encouragement Award, CTU
- Architecture Stack
- YOLOv8DeepSortPythonOpenCV
01 / Context & Problem Statement
Developed for Petrovietnam Ca Mau Fertilizer Corporation (PVCFC) during an onsite research project, this system automates inventory management in agricultural supply chains. The goal was to build a computer vision pipeline capable of detecting, tracking, and classifying fertilizer bags on moving conveyor belts under challenging industrial conditions.
02 / System Architecture & Pipeline Flow
Industrial Video Stream Ingestion
Captures high-vibration conveyor belt video feeds with motion blur and variable warehouse lighting.
Custom YOLOv8 Detection
Detects and classifies distinct PVCFC fertilizer bag SKUs in real time on each video frame.
DeepSort Kalman Tracking
Assigns persistent IDs across frames using Kalman filtering + appearance descriptors to prevent double-counting.
Virtual Line-Crossing Counter
Triggers inventory increments only when tracked bag centroids cross the calibrated inspection boundary.
03 / Engineering Approach
Trained a custom YOLOv8 model for high-speed object detection and classification. Integrated DeepSort for multi-object tracking to prevent double-counting as bags pass the camera sensor. The pipeline is written in Python using OpenCV for frame-by-frame processing and video ingestion, optimized to run with low latency on edge computing devices inside the manufacturing environment.
04 / Key Engineering Deliverables
- Custom-annotated industrial dataset of Ca Mau Fertilizer product lines
- YOLOv8 + DeepSort multi-object tracking resistant to occlusion
- Won Encouragement Award for practical industrial AI application
05 / Engineering Reflection
"Industrial AI is less about benchmark data and more about handling noise, dust, lighting changes, and camera vibration. Tuning YOLOv8 to achieve ~92% accuracy under variable physical conditions was a masterclass in robust dataset curation and real-time inference optimization."
