DualTake: Predicting Takeovers across Mobilities for Future Personalized Mobility Services
Zhaobo Zheng, Kumar Akash, Teruhisa Misu
TL;DR
This work tackles the problem of predicting driver takeovers across mobility types by leveraging traditional car-based behavioral data to anticipate takeovers in micro-mobility within a future hybrid mobility ecosystem. The authors introduce DualTake, a deep neural network augmented with transfer learning (TrAdaBoost) to enable cross-mobility driver monitoring, and validate it on a VR-based multimodal dataset of 142,321 samples collected from 48 participants. Their results show that a transfer-learning-enhanced model can achieve an accuracy of approximately 0.86 and an AUC of about 0.77 when predicting takeovers in micro-mobility from car data, outperforming a Random Forest baseline and a standalone DNN. This demonstrates the feasibility of universal driver-state understanding across mobilities and offers a path toward personalized, mobility-wide safety and user experience improvements in future transportation systems.
Abstract
A hybrid society is expected to emerge in the near future, with different mobilities interacting together, including cars, micro-mobilities, pedestrians, and robots. People may utilize multiple types of mobilities in their daily lives. As vehicle automation advances, driver modeling flourishes to provide personalized intelligent services. Thus, modeling drivers across mobilities would pave the road for future society mobility-as-a-service, and it is particularly interesting to predict driver behaviors in newer mobilities with traditional mobility data. In this work, we present takeover prediction on a micro-mobility, with car simulation data.The promising model performance demonstrates the feasibility of driver modeling across mobilities, as the first in the field.