Exploring single-head and multi-head CNN and LSTM-based models for road surface classification using on-board vehicle...

Area of Science:

• Engineering
• Computer Science
• Transportation Safety

Background:

• Road surface monitoring is critical for vehicle and pedestrian safety.
• Robust data acquisition and analysis are key to accurate road condition assessment.

Purpose of the Study:

• To classify road surface conditions using deep learning architectures.
• To evaluate Convolutional Neural Networks (CNNs) and CNNs combined with Long Short-Term Memory (LSTM) layers.
• To analyze data from Inertial Measurement Units (IMUs) on vehicle sprung and unsprung masses.

Main Methods:

• Implemented single- and multi-head deep learning models (CNNs, CNN+LSTM).
• Utilized acceleration and angular velocity data from IMUs at various vehicle positions.
• Performed hyperparameter tuning via grid search for optimal model configuration.

Main Results:

• CNN+LSTM models generally surpassed CNN-only models in performance.
• The top model, a single-head architecture with three IMUs, achieved a macro F1-score of 0.9338.
• Effectiveness of combining IMU data within a single-head architecture was demonstrated.

Conclusions:

• Deep learning, particularly CNN+LSTM, shows significant promise for road surface condition classification.
• Optimizing model architectures and expanding datasets can further enhance classification accuracy.
• This approach offers a pathway to improved road safety through advanced monitoring.