Analyzing Spatial Eye-Tracking Data of Teleoperators to Assess Workload in Automotive Fleet Management

Abstract

With the rise of automated vehicles, teleoperation plays a key role in ensuring safe and efficient drives, particularly in partially automated systems where human operators provide high-level commands. This research focuses on understanding how mental states, specifically cognitive workload, impact ocular behavior during teleoperation tasks using a visual interface. Data from a previously-conducted user study is analysed, where task difficulty and frequency were manipulated in a naturalistic setting. The data-driven approach stresses the use of spatial area-of-interest metrics and its evaluation in providing mental state insights. The results display workload to be a significant factor in influencing the selected AoI metrics, including fixation duration, fixation frequency, time-to-first fixation, visit frequency, dwell time, and stationary entropy. Moreover, the findings partially support that a high workload induces a tunneling effect, although modulated by task-related and interface factors. The influence of difficulty and frequency independently act on the AoI metrics, with the former eliciting a broader effect. The study also demonstrates that workload can be predicted using machine learning models, with binary workload and frequency predictions achieving high recall rates (above 85%), and difficulty prediction reaching a maximum of 75%. A 4-class workload classification has been attempted, too, with the best predictive model reaching a recall of 49%. These outcomes highlight the potential of AoI metrics for real-time workload assessment and detection in teleoperation, paving the way for intelligent interfaces that adapt to operator mental states.