Towards Increased Situational Awareness at Unstructured Work Zones: Analysis of Worker Behavioral Data Captured in VR-Based Micro Traffic Simulations

In 2020, the Federal Highway Administration (FHWA) reported 857 fatalities due to crashes at roadway work zones, 117 of which were construction workers. Current practices to protect the personal safety of workers, such as stationary sound and light-based work zone intrusion alarms, are frequently disregarded since their alarm characteristics (e.g., volume, duration, frequency) are not well suited to work zone environments. Such alarm devices are also rarely deployed at unstructured (i.e., mobile and short term) work zones. There is a need to identify optimal alarm characteristics that can effectively notify workers of incoming hazards at unstructured work zones without alarm fatigue. This paper is part of a larger research vision, where an integrated platform of virtual reality (VR), micro traffic simulations, and wearable sensors is used to capture workers’ physiological and behavioral response to alarms. This platform enabled analyzing behaviors during dangerous traffic scenarios that are not feasible to test in real world settings. The captured data will be used as a dataset to build reinforcement learning (RL) based calibration models to optimize the frequency, modality, and duration of alarms that workers are more attentive to. Towards this goal, this paper provides the results of initial data analysis on how workers responded to alarms sent to wearable devices with different modalities (e.g., sound, vibration), durations, and frequencies. Effect of different alarms on worker behavior and physiology has been measured with metrics that rely on workers’ positional changes, head movements and what they see in their field of views, heart rate changes, and acknowledgement of alarms on wearable interfaces. Results indicate that alarms with shorter durations and repeated less frequently can lead workers to react faster by detecting vehicles and acknowledging the wearable warning device. Findings from this study can serve as benchmarking data for the design of effective alarm systems for work zones and be utilized as a dataset for RL models for training agents for such systems.