An Adaptive Multi-Staged Forward Collision Warning System Using a Light Gradient Boosting Machine

Abstract

The existing forward collision warning (FCW) systems that adopt kinematic or perceptual parameters have some drawbacks in the warning performance because of poor adaptability to the users or ineffectiveness of the warnings. To solve the problems of adaptability, several FCW models have been proposed based on algorithms (machine learning, deep learning). However, there is a lack of consideration for the multi-staged warning to avoid an abrupt warning that may startle or distract the driver. In this study, a light gradient boosting machine (LGBM) was adopted to develop a multi-staged FCW. The proposed model was trained and evaluated on a platform based on a driving simulator by twenty drivers. Through Shapley Additive Explanations (SHAPs), the output of the proposed model was explained. Specifically, the front vehicle acceleration, time-to-collision (TTC), and relative speed were found to strongly affect the warning stages from the proposed model. To evaluate the utility and acceptability of the developed model, it was compared with three existing FCW models in terms of subjective and objective indicators. As a result, a trade-off was found between the utility and user acceptance. Additionally, the comparison study also indicated that the developed model outperformed other previous models due to not only the high accuracy but also the suitable trigger timing for each participant.