Real-Time Driver Fatigue Prediction System Based on IoT and Machine Learning

This project proposes an end-to-end IoT data pipeline and predictive intelligence system. Addressing issues like privacy infringement, light sensitivity, and high false alarm rates in traditional fatigue detection methods (e.g., camera monitoring), it integrates in-vehicle sensor data, driving behavior telemetry, and machine learning algorithms to achieve real-time prediction and early warning of driver fatigue and distraction risks. The system consists of three core components: a data science core layer, a FastAPI backend API, and a Streamlit visualization dashboard, with practical values in fleet safety management, intelligent driving assistance, and insurance industry applications.

Driver fatigue is one of the main causes of traffic accidents. Traditional camera monitoring methods have issues with privacy, light sensitivity, and false alarm rates. This project's innovative solution: integrate in-vehicle sensor data, driving behavior telemetry, and machine learning to build an end-to-end real-time fatigue prediction system, which not only detects the current state but also predicts risks in advance and issues warnings.

Architecture: Modular design with three core components: 1. Data science core layer (data generator, model training pipeline, model persistence); 2. Backend API service (built with FastAPI, providing real-time scoring, batch processing, model version management); 3. Visualization dashboard (built with Streamlit, supporting real-time monitoring, historical analysis, and warning configuration).

Technology Stack: Based on Python 3.13.5, relying on data processing and ML libraries like pandas, numpy, scikit-learn, xgboost; Web and visualization tools like FastAPI, uvicorn, Streamlit; Engineering tools like pydantic, joblib.

1. Multi-dimensional indicator fusion: Integrate four types of indicators to evaluate fatigue: PERCLOS (eye closure ratio >12% indicates high risk), yawning frequency, number of steering wheel micro-corrections, and continuous driving duration.

2. Intelligent data simulation engine: Generate 2000 realistic telemetry data entries, including progressive fatigue modeling, random noise injection, and expert rule annotation.

3. Ensemble learning model: Use random forest (100 trees, maximum depth of 8 layers, stratified sampling), which shows excellent classification performance on the test set.

1. Fleet safety management: 7x24 monitoring, risk behavior profiling, personalized training recommendations;

2. Intelligent driving assistance: Link with seat vibration, air conditioning cooling, and suggest rest at service areas;

3. Insurance industry: Risk level assessment, differentiated premiums, auxiliary accident liability determination.

1. Deep learning enhancement (processing time-series data to capture dynamic patterns);

2. Multi-modal fusion (integrating CAN bus, GPS, environmental sensors);

3. Edge computing deployment (optimizing models to support in-vehicle embedded devices);

4. Federated learning (aggregating multi-fleet data under privacy protection to improve models).

This open-source project demonstrates a complete machine learning engineering practice (data collection → model training → service deployment), provides practical fatigue detection functions, offers a reusable architecture template for IoT and AI integration applications, and serves as an excellent learning case and project starting point for developers in the fields of intelligent transportation and in-vehicle AI.