Extracting Features from Driving Scenarios for Driving Workload Level Classification - A Case Study of Transfer Learning

In the stage of automobile industry transition from SAE level “0,1” low autonomous through “2,3,4” human-in-the-loop and ultimately “5” fully autonomous driving, advanced driving monitor system is critical to understand the status, performance, and behavior of drivers for next-generation intelligent vehicles. By making necessary warnings or adjustments, they could operate collaboratively to ensure a safe and efficient traffic environment. The performance and behavior can be viewed as a reflection of the driver’s cognitive workload, which corresponds as well to the environment of their driving scenarios. In this study, image features extracted from driving scenarios, as well as additional environmental features were utilized to classify driving workload levels for different driving scenario video clips. As a continuing study of exploring transfer learning capability, two transfer learning approaches for feature extraction, image segmentation mask transfer approach and image-fixation map overlaid approach were compared and shown comparable results with a 0.910 AUC score and 0.918 AUC score respectively. Environmental information with easy accessibility also shown the effectiveness of contributing to the classification task as additional feature sources.