A GRU-Based Framework for Mobile Driver Behaviour Classification: Enhancements over LSTM Variants

Accurate driver behavior recognition is a key component in modern intelligent transportation systems, particularly for ride-sharing and fleet safety applications. Building on a previously proposed smartphone-based framework that used deep learning models to classify driving styles, this study revisits and extends that approach. We reimplemented three baseline models Long Short-Term Memory (LSTM), Convolutional LSTM (ConvLSTM), and CNN-LSTM exactly as described in the original system to ensure a fair evaluation under consistent conditions. In addition, we propose a new model based on the Gated Recurrent Unit (GRU) architecture, known for its simpler structure and efficiency. Using motion data captured from a smartphone’s accelerometer and gyroscope, we trained and compared all models for both multi-class and binary driver behavior classification. Experimental results show that the GRU model achieves the highest accuracy in the binary scenario, reaching 95.46%, and matches the best baseline performance for the multi-class task, while using fewer computational resources. This confirms that GRU is a practical and effective alternative for mobile-ready, real-time driver behavior monitoring.