Trajectory-level fog detection based on in-vehicle video camera with TensorFlow deep learning utilizing SHRP2...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Real-time weather information is vital for driver safety, especially in adverse conditions like fog.
• Existing systems for fog detection may lack affordability or real-time trajectory-level data capabilities.

Purpose of the Study:

• To develop an economical in-vehicle system for accurate, real-time fog detection.
• To provide trajectory-level weather data for enhancing road safety systems.

Main Methods:

• Utilized SHRP2 Naturalistic Driving Study (NDS) video data.
• Employed Deep Learning models: Deep Neural Network (DNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN).
• Trained models using Python and TensorFlow, comparing Adam and Gradient Descent optimizers across clear, distant fog, and near fog conditions.

Main Results:

• Convolutional Neural Network (CNN) achieved the highest accuracy (97% with Gradient Descent, 98% with Adam).
• The Adam optimizer generally improved prediction accuracy across all tested models compared to Gradient Descent.
• The system requires only a single video camera, indicating low implementation cost.

Conclusions:

• The proposed single-camera, deep learning-based fog detection method is an affordable solution for real-time weather data collection.
• This technology can significantly enhance weather-based Variable Speed Limit (VSL) and Advanced Traveler Information Systems (ATIS).