Feature-Transfer based Driving Control Model Guided by Driver Attention

Many intelligent vehicular and transportation applications require the ability to predict vehicle speed. However, it is challenging to accurately predict on-road vehicle speed since individual vehicle speed could be impacted (in either a stochastic or deterministic manner) by numerous traffic-driver-road system related variables, e.g. route type, traffic condition and driver’s behavior. Nowadays, machine learning techniques are adopted to explore the potential correlation between the vehicle speed and its primary factors using the historical driving data, which makes vehicle speed prediction more reachable. In this paper, a novel data-driven method, Feature-transfer Driving Control Model (FDCN) is proposed for accurate prediction of short-term individual vehicle speed from vehicle signals and front view video frames. To bridge the gap between two modalities, a cross-modal feature fusion (CFF) module is custom-designed, which aggregates the extracted vehicle dynamics feature and driver gaze feature effectively to produce a cross-modal representation. Besides, a feature-transfer (FT) module is developed to distill the driver-gaze-related information from cross-modal representation to the vehicle dynamics feature branch efficiently. Finally, two driving data set are employed to evaluate the prediction performance. As a result, the proposed FDCN outperforms other advanced models in terms of RMSE. Particularly, this work demonstrates the effectiveness of FDCN in pedestrian-involved scenarios.