🧥 Staff Uniform SOP Compliance System

An end-to-end computer vision system that automatically detects staff members in CCTV footage and verifies whether they are wearing the correct uniform top.
The system combines YOLO-based detection with a fine-tuned EfficientNet classifier, allowing for robust and automated daily compliance scanning.

---

🎯 Real-World Use Case

Objective: Automatically assess uniform compliance once per day using surveillance cameras (e.g., Camera 05 & Camera 06).

Workflow Summary

1. Detect people in the frame using YOLO.
2. Detect the top-wear region (shirt/vest) from each person using a fine-tuned YOLO trained on a custom dataset.
3. Classify the top-wear as Compliant (Polo Compliant / Vest Compliant) or Non-compliant using a fine-tuned EfficientNet-B0 model.
4. Generate visual inference results with colored bounding boxes.

---

🧩 System Architecture

[ CCTV Frame ]
     │
     ▼
 ┌─────────────────────────────┐
 │ YOLOv11n (Person Detector)  │
 └─────────────────────────────┘
     │ Cropped Persons
     ▼
 ┌──────────────────────────────────────────┐
 │ Fine-Tuned YOLOv11n (Top-wear Detector)  │
 └──────────────────────────────────────────┘
     │ Cropped Shirts
     ▼
 ┌─────────────────────────────┐
 │ EfficientNet-B0 Classifier  │
 │ (Compliant / Non-compliant) │
 └─────────────────────────────┘
     │
     ▼
[ Annotated Inference Results ]

---

⚙️ Project Components

Component	Description
YOLO (Person Detector)	Detects staff members in raw CCTV images.
YOLO (Top-wear Detector)	Detects shirt/vest region for each detected person.
EfficientNet-B0 Classifier	Classifies each detected shirt as polo_compliant, vest_compliant, or non-compliant.
CLIP (Optional)	Used for intelligent dataset filtering and visual data sorting during preparation.

---

🧰 Directory Structure

.
├── cctv5/                          # CCTV 05 input images
├── cctv6/                          # CCTV 06 input images
│
├── detected_shirts/                # OUTPUT: Cropped shirt regions
├── inference_results/              # OUTPUT: Annotated results
│
├── models/
│   ├── yolo11n.pt                  # Pretrained YOLO model (person)
│   ├── top_wear_detector.pt        # Custom YOLO model (shirt)
│   ├── best_ppe_classifier.pth     # Fine-tuned EfficientNet model
│   └── class_to_idx.json           # Class index map
│
├── scripts/
│   ├── detect_person.py
│   ├── detect_shirt.py
│   ├── classify_detected.py
│   ├── prepare_dataset.py
│   ├── fine_tune_model.py
│   └── main.py                     # Integrated inference pipeline
│
├── requirements.txt
└── README.md

---

🚀 How to Run

1️⃣ Install Dependencies

pip install -r requirements.txt

---

2️⃣ Prepare Dataset

Use the script below to organize and augment your dataset for training.

python scripts/prepare_dataset.py

This will automatically:

• Split data into train/validation sets.
• Apply data augmentation (rotation, flip, brightness, contrast).
• Ensure balanced data across compliant and non-compliant categories.

---

3️⃣ Fine-Tune EfficientNet-B0

Run the following command to fine-tune your classification model:

python scripts/fine_tune_model.py

This will:

• Load EfficientNet-B0
• Train on your prepared dataset
• Save model as best_ppe_classifier.pth
• Generate class_to_idx.json for class mapping

---

4️⃣ Run Full Inference Pipeline

Execute:

python scripts/main.py

This script will:

• Load YOLOv11n and EfficientNet models.
• Process all CCTV images from cctv5/ and cctv6/.
• Detect persons and their shirts.
• Classify detected shirts into compliance categories.
• Annotate and save inference results.

---

🎨 Visualization Legend

Label	Meaning	Bounding Box Color
polo_compliant	Staff wearing approved polo uniform	🟡 Yellow
vest_compliant	Staff wearing approved vest uniform	🟦 Cyan
non_compliant	Staff not wearing correct top wear	🔴 Red

---

⚡ Configuration Overview (main.py)

Variable	Description
PERSON_MODEL_PATH	YOLO model for person detection (yolo11n.pt)
SHIRT_MODEL_PATH	YOLO model for shirt detection (top_wear_detector.pt)
FINETUNED_MODEL_PATH	EfficientNet model path
CLASS_MAP_PATH	Path to class-to-index mapping file
SOURCES_TO_PROCESS	List of CCTV folders to process
CONFIDENCE_THRESHOLD	Minimum confidence for classification
COLOR_MAP	RGB color codes for visualization

---

🧠 Model Summary

Model	Framework	Purpose	Notes
YOLOv11n	Ultralytics	Person detection	Lightweight and real-time capable
YOLOv11n (custom)	Ultralytics	Top-wear detection	Fine-tuned on staff torso dataset
EfficientNet-B0	PyTorch	Classification	Fine-tuned for compliant vs non-compliant

---

🧪 Example Inference Output

Example of annotated CCTV frame with compliance results:

[Person Box] → "polo_compliant"
[Person Box] → "Vest_compliant"
[Person Box] → "non_compliant"

Each processed image is automatically saved in:

• inference_results/ → Annotated compliance images
• detected_shirts/ → Cropped top-wear regions

---

📅 Deployment Note

This system is designed for daily uniform compliance scans (e.g., once per day per camera).
It can be optionally integrated with a face recognition attendance system to correlate:

• Detected faces (identity)
• Uniform compliance status

---