Factory Counting System MVP

A real-time computer vision system for counting chocolates and bottles on conveyor belts using YOLO object detection and tracking.

📋 Table of Contents

• Overview
• Features
• Project Structure
• Requirements
• Installation
• Usage
• Model Training
• Configuration
• Results
• Troubleshooting

Overview

This project implements an automated counting system for factory conveyor belt. Using YOLO11 object detection and tracking, currently the system counts chocolates and bottles as they cross a user-defined line in real-time.in future carious other classes will be added .Trained model is not added in this repo . One can use their own trained model change the model paths in pharms.yaml file .

Key Capabilities

• Real-time counting: Track and count objects crossing a defined line
• Bi-directional tracking: Separate counts for IN and OUT directions
• Multi-class detection: Chocolates and bottles
• Multi-camera support: Process multiple camera feeds simultaneously
• Visual feedback: Neon-style bounding boxes with color-coded states
• MLflow integration: Track training experiments and model versions

Features

• 🎯 Accurate Detection: YOLO11 model trained on factory conveyor data
• 📊 Class-specific Counting: Tracks counts per object type (chocolate, bottle)
• 🎨 Neon Visual Effects: Modern corner-box visualization with glow effects
• 🎥 Flexible Input: Support for video files, RTSP streams, and webcams
• 📈 Experiment Tracking: MLflow integration for model versioning
• 🔄 Persistent Tracking: Object IDs maintained across frames
• ⚡ Real-time Processing: Efficient multi-threaded architecture

📁 Project Structure

factory_chocolat_counting/
│
├── src/
│   ├── data_ingestion.py              # Download datasets from Roboflow
│   ├── model_building.py              # Train YOLO model with MLflow
│   ├── postprocessing_bisunesslogic.py           # Single-camera counting
│   └── postprocessing_bisunesslogic_multicameraprocessing.py  # Multi-camera counting
│
├── parms.yaml                         # Training configuration
├── requriments.txt                    # Python dependencies

🔧 Requirements

• Python 3.8.11
• CUDA-capable GPU (recommended)

📦 Installation

1. Clone the repository

git clone [repository_url]
cd factory_chocolat_counting

2. Create virtual environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

# install from requirements file
pip install -r requriments.txt

🚀 Usage

Single Camera Counting

For single webcam, RTSP stream, or video file:

python src/postprocessing_bisunesslogic.py

Interactive setup:

1. Choose input source (video file or camera)
2. Select counting line: Click 2 points to define the line
3. Watch real-time counting with visual feedback

Controls:

• q: Quit
• ENTER: Confirm line selection
• ESC: Exit during line selection

Multi-Camera Counting

For multiple simultaneous camera feeds:

python src/postprocessing_bisunesslogic_multicameraprocessing.py

Setup:

1. Enter camera indices (e.g., 0 1 for two cameras)
2. Draw counting line for each camera
3. System processes all feeds in parallel

Counting Logic

The system uses line-crossing detection:

• Red bounding box: Object detected, not crossed yet
• Yellow flash: Object just crossed the line (4 frames)
• Green bounding box: Object already counted
• Counts displayed: [CLASS_NAME] IN: X OUT: Y

Line-crossing algorithm: Detects which side of the line an object's center point is on and tracks direction changes.

🔬 Model Training

Train Your Own Model

1. Configure parameters in parms.yaml, please modify the parameters according to your project setup:

model_path: yolo11n.pt
data_yaml_path:Project-3/data.yaml
epochs: 100
batch: 4
imgsz: 640
device: 0  # GPU device
experiment_name: "choclate_identification_yolo_11"
mlflow_tracking_uri: "http://12.0.0.1:9000"

2. Set up MLflow tracking server:

mlflow server 

3. Run training:

python src/model_building.py

Trained models are automatically logged to MLflow with metrics and artifacts.

4. View results in MLflow UI:

# MLflow UI should be accessible at http://localhost:5000
# View experiments, compare runs, and download models

⚙️ Configuration

Training Parameters (parms.yaml)

• model_path: Pre-trained YOLO model (e.g., yolo11n.pt)
• data_yaml_path: Dataset configuration file
• epochs: Number of training epochs (default: 100)
• batch: Batch size (default: 4)
• imgsz: Image size (default: 640)
• device: GPU/CPU device (0 for GPU, cpu for CPU)
• experiment_name: MLflow experiment name
• mlflow_tracking_uri: MLflow server URL

Detection Parameters (in scripts)

• Confidence threshold: 0.7 (hardcoded)
• Classes: [0, 1] (chocolate=0, bottle=1)
• Tracking: Persistent tracking enabled (persist=True)
• Flash duration: 4 frames after crossing

📊 Results

Model Performance

• Classes: Chocolate, Bottle
• Dataset v3: 780 train / 75 valid / 37 test images
• Dataset v2: 570 train / 55 valid / 27 test images
• Best Model: runs/detect/train5/weights/best.pt
• Format: YOLO8/OBB format

Counting Features

• Bi-directional counting: Accurate IN/OUT detection using line-crossing algorithm
• Multi-threaded: Efficient processing of multiple camera feeds
• Visual feedback: Real-time neon-style corner boxes
• Frame persistence: Objects tracked across frames with unique IDs
• Per-class tracking: Separate counts for each object type

🎥 Visual Feedback

The system displays:

• Neon corner boxes: Modern corner-only visualization with glow effects
• Center tracking: Cyan circle at object center
• Class labels: Object type and confidence score
• Counting line: Red line showing detection boundary
• Count displays: Real-time IN/OUT counts per class

Color coding:

• 🔴 Red: Newly detected object
• 🟡 Yellow: Just crossed (flash effect)
• 🟢 Green: Already counted

sample outputs use case 1: cloclate factory counting

case 2: bottle factory counting

🔍 Dataset Information

Dataset Version 3

• Classes: 2 (bottle, chocolate)
• Total images: 892
• Format: YOLO11
• Source: Roboflow factory conveyor belt footage

datasets validated with matching image/label pairs. The dataset can be accessed via this url : https://universe.roboflow.com/monitorshrimps/industrial_dataset-xzhwj

📝 License

See LICENSE file for details.

🐛 Troubleshooting

Issue: Camera not opening

• Check camera permissions
• Verify camera index is correct
• Try different indices (0, 1, 2...)

Issue: Model not found

• Verify best.pt exists in runs/detect/train5/weights/
• Update path in scripts if moved

Issue: Poor detection accuracy

• Adjust confidence threshold in scripts
• Retrain with more epochs
• Use more diverse training data

Issue: Multi-camera setup fails

• Ensure camera indices are correct
• Check if cameras are already in use
• Verify sufficient USB bandwidth

---

Built with using Ultralytics YOLO, OpenCV, and MLflow