Identification of Driver Cognitive Workload Using Support Vector Machines with Driving Performance, Physiology and Eye Movement in a Driving Simulator Joonwoo Son 1,# , Hosang Oh 1 , and Myoungouk Park 1 1 DGIST HumanLab, 50-1 Sang-Ri, Hyeonpung-Myeon, Dalseong-Gun, Daegu, South Korea, 711-873 # Corresponding Author / E-mail: json@dgist.ac.kr, TEL: +82-53-785-4740, FAX: +82-53-785-4759 KEYWORDS: Driver state estimation, Cognitive workload, Support vector machines, Intelligent vehicle, Driving simulator This paper suggests experimental approaches for identifying driver’s cognitive workload using support vector machines (SVMs) with driving performance, physiological response and eye movement data. In order to construct a classification model for detecting high cognitive workload condition, driving simulation experiments were conducted. For the experiments, 30 participants (15 younger males in the 25-35 age range (M = 27.9, SD = 3.13) and 15 older males in the 60-69 (M = 63.2, SD = 1.74)) drove a simulated highway in a fixed-base driving simulator. While driving through 37 km of straight highway, participants conducted three levels of cognitive secondary tasks, i.e. an auditory delayed digit recall task, at specified segments for 10 minutes and their driving performance, physiological response and eye movement data were collected. In this study, the model performances with different combination of measures were assessed with the nested cross-validation method. As a result, it was demonstrated that the proposed SVM models were able to identify driver’s cognitive workload with high accuracy. The best performance was achieved with a combination of the standard deviation of lane position (SDLP), physiology and gaze information. The best model obtained 89.0% accuracy, sensitivity of 86.4% and specificity of 91.7%. 1. Introduction Driver inattention causes a significant problem for road traffic safety, because it degrades driving performance and situational awareness. According to car accident statistics, between 25% and 78% of crashes are caused by driver inattention. 1 Drivers’ cognitive workload, when it is too low (e.g. fatigue or drowsiness) or too high (e.g. stress or multiple tasks), is related to driver inattention and accident. 2,3 Thus, a proper identification of a driver’s workload and spare capacity is one of the promising approaches to designing an adaptive automotive user interface for reducing driver distraction. 4 By monitoring driver’s workload, the adaptive interface can provide timely and targeted information when the driver has the spare capacity. According to the final report of the Driver Workload Metric Project 5 in the United States, driver workload is defined as the competition in driver resources between the driving task and a concurrent subsidiary task, occurring over the task’s duration, as manifested in degraded lane keeping, longitudinal control, object-and-event detection, or eye glance behavior. Two major types of driving workload are visual and cognitive workload. Visual demand is straightforward, but cognitive workload is difficult to measure directly because it is essentially internal to the driver. 6 Nevertheless, there have been efforts to measure cognitive workload using subjective measures, 7,8 physiological measures, 9-12 eye movement measures, 13,14 and driving performance measures. 10,14 Among those measures, driving performance measures can detect the cognitive workload using easy and less expensive methods through readily available in-vehicle information. 15,16 However, driving performance measures are known to have limitations compared to others due to their reflected small changes with respect to the cognitive workload changes. 6,11,15 On the other hand, physiological measures have been proposed as useful metrics for assessing workload. Mehler et al. found that a near linear increase in heart rate and skin conductance appeared as the workload levels increase. 11 Then, Son et al. 17 demonstrated that the combination of performance and physiological data can be effectively used for estimating cognitive workload, but the data sets in their study