Cross-subject driver status detection from physiological signals based on hybrid feature selection and transfer learning

• Published in: Expert systems with applications Vol. 137; pp. 266 - 280
• Main Authors: Chen, Lan-lan, Zhang, Ao, Lou, Xiao-guang
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 15.12.2019; Elsevier BV
• Subjects: Algorithms; Computer simulation; Cross-subject feature evaluation; Driver status; Drivers; Driving; Evaluation; Hybrid feature selection; Knowledge management; Machine learning; On-line systems; Physiological signals; Physiology; Traffic accidents; Transfer learning; Cross-subject feature evaluation; Driver status; Physiological signals; Transfer learning; Hybrid feature selection
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2019.02.005

•Cross-subject learning framework for driver status detection was proposed.•Features were estimated from both class separation and domain fusion perspectives.•Hybrid feature selection method based on multiple filter algorithms was employed.•Several domain adaptation transfer learning techniques were adopted and compared.•Both simulated and real drive datasets verified the effectiveness of proposed method. It is a challenging and rewarding work for monitoring driving status in daily commute, which is favorable for declining the occurrence of traffic accidents and promoting driver's health. One big challenge that restricts this kind of research from real-life applications is the robustness and transfer ability of learning methods that can effectively tackle individual difference. Drawing knowledge from others through transfer learning could boost detection performance of a new driver. The present study aims to develop an efficient cross-subject transfer learning framework for driving status detection based on physiological signals. To grasp what part of knowledge was appropriate for transferring, cross-subject feature evaluation was used to measure feature quality. Then based on the evaluation score, several filtering algorithms were combined to search for better feature subsets that were not only helpful for later classification tasks but also robust to the individual difference. Finally, the framework based on hybrid feature selection and efficient transfer classifier was validated using simulated and real driving datasets. Our experimental results revealed that the proposed algorithm could achieve high recognition accuracy and good transferability among individuals, which could increase the scope of application of physiological data for drive status detection during daily life, as it alleviated the need of subject specific pilot data for assessing the physiological characteristics across subjects. This scheme can be further developed into an online warning and assistant system in vehicles helping to early detect driver's unfavorable status, better manage their negative emotion and decrease the occurrence of traffic accidents.