Development and Evaluation of a Computational Human Performance Model of In-vehicle Manual and Speech Interactions

Abstract

Usability evaluation traditionally relies on costly and time-consuming human-subject experiments, which typically involve developing physical prototypes, designing usability experiment, and recruiting human subjects. To minimize the limitations of human-subject experiments, computational human performance models can be used as an alternative. Human performance models generate digital simulations of human performance and examine the underlying psychological and physiological mechanisms to help understand and predict human performance. A variety of in-vehicle information systems (IVISs) using advanced automotive technologies have been developed to improve driver interactions with the in-vehicle systems. Numerous studies have used human subjects to evaluate in-vehicle human-system interactions; however, there are few modeling studies to estimate and simulate human performance, especially in in-vehicle manual and speech interactions. This paper presents a computational human performance modeling study for a usability test of IVISs using manual and speech interactions. Specifically, the model was aimed to generate digital simulations of human performance for a driver seat adjustment task to decrease the comfort level of a part of driver seat (i.e., the lower lumbar), using three different IVIS controls: direct-manual, indirect-manual, and voice controls. The direct-manual control is an input method to press buttons on the touchscreen display located on the center stack in the vehicle. The indirect-manual control is to press physical buttons mounted on the steering wheel to control a small display in the dashboard-cluster, which requires confirming visual feedback on the cluster display located on the dashboard. The voice control is to say a voice command, “ deflate lower lumbar” through an in-vehicle speaker. The model was developed to estimate task completion time and workload for the driver seat adjustment task, using the Queueing Network cognitive architecture (Liu, Feyen, & Tsimhoni, 2006). Processing times in the model were recorded every 50 msec and used as the estimates of task completion time. The estimated workload was measured by percentage utilization of servers used in the architecture. After the model was developed, the model was evaluated using an empirical data set of thirty-five human subjects from Chen, Tonshal, Rankin, & Feng (2016), in which the task completion times for the driver seat adjustment task using commercial in-vehicle systems (i.e., SYNC with MyFord Touch) were recorded. Driver workload was measured by NASA’s task load index (TLX). The average of the values from the NASA-TLX’s six categories was used to compare to the model’s estimated workload. The model produced results similar to actual human performance (i.e., task completion time, workload). The real-world engineering example presented in this study contributes to the literature of computational human performance modeling research.