Project XCEED – Real-Time Seat Belt Detection System

Maruti Suzuki India Limited (MSIL) — Project Xceed Internship Program
Edge AI · YOLOv8 · Raspberry Pi 5 · Fully Offline · Real-Time

Project XCEED is a real-time seat belt compliance monitoring system developed as part of the Maruti Suzuki India Limited internship program. The system uses a YOLOv8-based computer vision pipeline deployed on a Raspberry Pi 5 to detect seat belt usage and identify unsafe occupant behaviour in real time.

The solution operates fully offline, is optimised for edge deployment, and provides both a software monitoring dashboard and GPIO-based hardware alerts. It was validated across multiple lighting conditions, clothing contrasts, and seat belt misuse scenarios.

Build a camera-based seat belt monitoring system capable of:

• Detecting seat belt compliance in real time
• Identifying non-compliance scenarios including misuse and evasion
• Triggering immediate warnings upon violation detection
• Running on low-compute edge hardware without GPU acceleration
• Operating without internet connectivity during inference
• Remaining robust across varied lighting conditions and pose variations

• Raspberry Pi 5 (2 GB) — no cloud dependency
• ONNX Runtime CPU inference at 18–25 FPS
• Fully offline; zero network calls during operation
• GPIO-controlled LED and buzzer hardware alerts

Each violation class triggers a distinct buzzer pattern so the alert type is identifiable without looking at the dashboard:

Alert stability is managed by hysteresis logic — Frame Buffer (FBS = 3) to activate and Clean Buffer (CBS = 16) to deactivate — eliminating flickering from single-frame misclassifications.

The Streamlit dashboard provides:

• Live detection class and confidence score
• Active alert state
• Violation log with timestamps
• Session statistics (total inferences, alert events, per-class breakdown)
• Database-backed traceability

All detection events are persisted to xceed.db (SQLite):

xceed.db is excluded from version control and is generated at runtime.

During validation, the system logged 28,159 inference records and 10,641 alert events across all test conditions.

project-xceed-msil/
│
├── ai/                              # AI subsystem
│   ├── detect.py                    # Real-time inference loop (main runtime)
│   ├── train_yolo.py                # YOLOv8 training script
│   ├── extract_frames.py            # Video-to-frame dataset extractor
│   ├── capture_custom.py            # Pi Camera data collection utility
│   ├── camera_test.py               # Camera connectivity test
│   └── live_preview.py              # Live camera preview utility
│
├── backend/
│   └── main.py                      # FastAPI REST API — detection logging and dashboard data
│
├── frontend/
│   └── dashboard.py                 # Streamlit monitoring dashboard
│
├── hardware/
│   └── gpio_alert.py                # GPIO LED and buzzer controller
│
├── datasets/                        # Active training datasets
│   ├── classes.txt                  # Class label definitions
│   ├── final_dataset/               # Combined RGB dataset (production training)
│   │   ├── data.yaml                # ✓ tracked
│   │   ├── images/                  # ✗ excluded — large binary assets
│   │   └── labels/                  # ✗ excluded — large binary assets
│   ├── final_dataset_ir_raw/        # Raw infrared dataset
│   │   ├── data.yaml                # ✓ tracked
│   │   ├── images/                  # ✗ excluded
│   │   └── labels/                  # ✗ excluded
│   └── final_dataset_ir_clahe/      # CLAHE-enhanced infrared dataset
│       ├── data.yaml                # ✓ tracked
│       ├── images/                  # ✗ excluded
│       └── labels/                  # ✗ excluded
│
├── models/                          # Model artifacts
│   ├── best320.onnx                 # ✓ tracked — production deployment model (320×320)
│   ├── best_final.pt                # ✓ tracked — final PyTorch checkpoint (for retraining)
│   └── yolov8n.pt                   # ✓ tracked — base YOLOv8n weights
│
├── scripts/                         # Dataset preparation utilities
│   ├── merge_datasets.py            # Merge RGB dataset sources
│   ├── merge_datasets_ir_raw.py     # Merge raw IR dataset sources
│   ├── merge_datasets_ir_clahe.py   # Merge CLAHE IR dataset sources
│   ├── create_clahe_dataset.py      # Apply CLAHE to raw IR frames
│   └── validate_dataset.py          # Verify annotation integrity and class balance
│
├── demo_recordings/                 # Per-scenario validation recordings
│   ├── 01_bright_white/             # Daylight, white shirt — all four classes
│   │   ├── proper_belt/             # ✗ video files excluded
│   │   ├── no_belt/
│   │   ├── clipped_behind/
│   │   └── decoy/
│   ├── 02_bright_black/             # Daylight, black shirt — all four classes
│   ├── 03_dim_white/                # Active cabin, white shirt — all four classes
│   ├── 04_dim_black/                # Active cabin, black shirt — all four classes
│   ├── 05_full_demo/
│   │   └── output.mp4               # ✗ excluded
│   └── 06_offline_proof/
│       └── output.mp4               # ✗ excluded
│
├── media/                           # Additional media assets
│   ├── deployment_tests/
│   │   ├── output_master.mp4        # ✗ excluded
│   │   └── output320_detect.mp4     # ✗ excluded
│   └── model_comparison/
│       ├── best320.mp4              # ✗ excluded
│       ├── ir_raw.mp4               # ✗ excluded
│       └── ir_clahe.mp4             # ✗ excluded
│
├── docs/                            # Documentation assets — all tracked
│   ├── System Architecture.jpeg
│   ├── Hardware Architecture.jpeg
│   ├── Alert Logic Flowchart.jpeg
│   ├── MODEL_DEVELOPMENT.jpg
│   ├── model_comparison.jpg
│   ├── dashboard.jpg
│   ├── hardware_setup.jpg
│   ├── gpio_wiring.jpg
│   └── ir_illuminator.jpg
│
├── reports/                         # Submitted project reports — all tracked
│   ├── Initial Plan Report — Project Xceed.pdf
│   ├── Mid_Progress_Report_FINAL.pdf
│   └── DAKSHAYANI_SHARMA_XCEED.pdf
│
├── archive/                         # Legacy development artifacts
│   ├── dataset.yaml
│   ├── datasets/                    # ✗ excluded — raw collected data before merging
│   ├── training_history/            # ✗ excluded — per-phase YOLOv8 run outputs
│   └── videos/                      # ✗ excluded — source videos for dataset extraction
│
├── master.sh                        # System orchestration entry point
├── requirements.txt                 # Python dependency specification
└── README.md

Note on excluded assets:
Dataset images/labels, video recordings, and intermediate training weights are excluded from this repository via .gitignore due to file size. best320.onnx, best_final.pt, and yolov8n.pt are tracked. All other excluded assets are available in the project submission package provided to MSIL.

The Pi Camera V2 NoIR was selected because it lacks an infrared cut filter, enabling imaging under near-zero-lux conditions when paired with the IR illuminator.

git clone https://github.com/dakshayani-codes/project-xceed-msil.git
cd project-xceed-msil

python3 -m venv xceed-env
source xceed-env/bin/activate

pip install -r requirements.txt

The ONNX deployment model is tracked in this repository at:

models/best320.onnx

If it is missing after cloning (e.g. due to a shallow clone or separate distribution), obtain it from the project submission package and place it at the path above before proceeding.

bash master.sh

master.sh automatically:

• Activates the virtual environment
• Starts the FastAPI backend on port 8000
• Starts the Streamlit dashboard on port 8501
• Launches the real-time inference pipeline
• Initialises the GPIO hardware alert subsystem

From the Raspberry Pi:

http://localhost:8501

From another device on the same network:

http://<raspberry-pi-ip>:8501

Find your Pi's IP address by running on the Pi:

hostname -I

Trained on a curated public seat belt dataset. Established the detection baseline and exposed deployment limitations: class imbalance, clothing-colour sensitivity, and weak clipped-behind recall.

Dataset expanded with custom images captured using the Pi Camera V2 NoIR, manually photographed scenarios, and video-extracted frames. Targeted coverage of under-represented conditions: no-belt on white shirt, clipped-behind geometry, and decoy objects common in the Indian market (dupattas, lanyards, bag straps). Produced best_final.pt, exported to best320.onnx.

best_final.pt is tracked in this repository. best320.onnx is the tracked production ONNX artifact.

Two infrared models were obtained through transfer learning by fine-tuning the final RGB model (best_final.pt) on infrared datasets:

• best_ir_raw.onnx — fine-tuned on raw infrared imagery
• best_ir_clahe.onnx — fine-tuned on CLAHE-enhanced infrared imagery

Both models were trained for 15 additional epochs and evaluated for passive-cabin operation under near-zero-lux conditions. CLAHE improved local contrast at belt-torso boundaries and outperformed the raw IR model. Both IR models remain experimental; best320.onnx is the production deployment model. IR ONNX exports are not tracked in this repository.

best320.onnx was selected for deployment because it provides the best real-world balance of throughput, robustness across lighting conditions, and operational reliability. The higher mAP values of the IR models reflect in-distribution validation only; real-world IR performance is lower due to domain shift from the RGB training distribution.

The system was tested across the following conditions:

Decoy objects tested include bag straps, lanyards, cables, and dupattas.

A representative end-to-end demonstration video (demo_recordings/05_full_demo/output.mp4) is included in this repository. Complete scenario-wise validation recordings are available in the project submission package provided to MSIL.

The project can be reproduced using:

pip install -r requirements.txt
bash master.sh

requirements.txt defines all Python dependencies required for deployment. master.sh orchestrates backend startup, dashboard initialisation, inference execution, and hardware alert activation. These two files serve distinct and complementary roles: requirements.txt enables environment portability; master.sh enables operational orchestration.

All experiments, datasets, training history, and deployment artifacts are documented within this repository and the accompanying project reports in reports/.

• Very low-contrast combinations (black belt on black shirt) remain challenging in dim light
• Dupattas and diagonal garments occasionally misclassify as proper_belt
• clipped_behind detection is geometry-dependent and sensitive to camera angle
• Third-row passenger monitoring was not validated (single-camera field of view)
• Infrared models show residual domain shift compared to the RGB deployment model

The docs/ folder contains all tracked visual assets:

Video demonstrations are excluded from version control. See the project submission package for full recordings.

Dakshayani Sharma
Project Xceed Program
Maruti Suzuki India Limited (MSIL), 2026