Automatic driver stress level classification using multimodal deep learning

© 2019 Stress has been identified as one of the contributing factors to vehicle crashes which create a significant cost in terms of loss of life and productivity for governments and societies. Motivated by the need to address the significant costs of driver stress, it is essential to build a practical system that can detect drivers’ stress levels in real time with high accuracy. A driver stress detection model often requires data from different modalities, including ECG signals, vehicle data (e.g., steering wheel, brake pedal) and contextual data (e.g., weather conditions and other ambient factors). Most of the current works use traditional machine learning techniques to fuse multimodal data at different levels (e.g., feature level) to classify drivers’ stress levels. Although traditional multimodal fusion models are beneficial for driver stress detection, they inherently have some critical limitations (e.g., ignore non-linear correlation across modalities) that may hinder the development of a reliable and accurate model. To overcome the limitations of traditional multimodal fusion, this paper proposes a framework based on adopting deep learning techniques for driver stress classification captured by multimodal data. Specifically, we propose a multimodal fusion model based on convolutional neural networks (CNN) and long short-term memory (LSTM) to fuse the ECG, vehicle data and contextual data to jointly learn the highly correlated representation across modalities, after learning each modality, with a single deep network. To validate the effectiveness of the proposed model, we perform experiments on our dataset collected using an advanced driving simulator. In this paper, we present a multi-modal system based on the adoption of deep learning techniques to improve the performance of driver stress classification. The results show that the proposed model outperforms model built using the traditional machine learning techniques based on handcrafted features (average accuracy: 92.8%, sensitivity: 94.13%, specificity: 97.37% and precision: 95.00%).