以深度學習演算法預測國道客運高速公路ADAS警示事件

Predicting ADAS Warning Events in Intercity Buses on Highways using Deep Learning Algorithms

過去較缺乏針對公車客運業者先進駕駛輔助系統（Advanced Driver Assistance Systems，ADAS）警示事件預測之研究。本研究使用臺灣H國道客運業者之自然駕駛資料，將逐秒資料合併為應變數與自變數時間窗，比較不同時間聚合方法與深度學習演算法在預測ADAS警示事件之優劣，並利用SHAP（SHapley Additive exPlanations）值分析變數間關係。本研究發現長短期記憶網路（Long Short-term Memory，LSTM）及遞迴神經網路（Recurrent Neural Network，RNN）之預測結果較佳，也發現應變數時間窗為5秒，自變數時間窗為往前10秒乃最佳之時間聚合方式。時間窗內車速最大值、車速標準差、左右方向燈及是否行駛於都會區與警示事件呈正相關；而時間窗內轉向角度最大值、前一期有無警示、下雨與衝度則呈負相關。本文最後討論如何將所發展之ADAS警示模式應用於車隊安全管理。

Few studies focused on the analysis of prediction of warning events related to Advanced Driver Assistance Systems (ADAS) for bus operators. This paper used naturalistic driving data provided by a Taiwan freeway bus operator and aggregated per-second data into dependent and independent variable time windows. The study compared the performance of different time aggregation types and deep learning algorithms in predicting ADAS warning events. Additionally, the relationships between variables were examined using SHAP (SHapley Additive exPlanations) value. This study demonstrated that Long Short-term Memory (LSTM) and Recurrent Neural Network (RNN) algorithms were superior to other algorithms. Moreover, the analytical results showed that the optimal time aggregation type occurred when the dependent variable time window was set to 5 seconds and the independent variable time window was set to 10 seconds prior. The max value of speed, the standard deviation of speed, the dummy variable of use of signal light, and the dummy variable of driving in metropolitan area within the independent variable time window had positive relationships with warning events. On the other hand, the max value of steering angle, the presence of warning event(s) in the previous time window, raining, and jerk were negatively associated with warning events. Finally, this study discussed how the proposed ADAS warning events model can be applied to fleet safety management.