以ADAS警示為基礎之國道大客車高風險駕駛行為辨識與預測

IDENTIFYING AND PREDICTING RISKY DRIVING BEHAVIORS IN HIGHWAY BUS DRIVERS BASED ON ADAS WARNINGS

先進駕駛輔助系統（Advanced Driver Assistance System，ADAS）已是大客車必備的車輛設備，然而ADAS警示並非完全準確，頻繁假警示可能會導致駕駛習慣性忽略警示，反而增加行車危險。為解決此問題，本研究發展一套以ADAS警示為基礎之高風險駕駛行為辨識流程，首先定義及標示行車異常事件作為ADAS警示之真值，接著利用二元羅吉特迴歸（binary logit regression）篩選與事件異常程度顯著相關變數，並以增幅式迴歸樹（boosted regression tree）辨識連續變數與異常事件的非線性關係，據以制定連續變數門檻，最後計算各變數組合下之車內外情境對應之事件異常化機率，以評估並找出高風險駕駛情境。本研究以國道客運576個縱向警示 （例如高速前車踫撞） 及425個橫向偏移警示之資料驗證所提方法之可行性，分析結果發現車內分心行為（例如拿東西、用手機等）為橫向偏移高風險駕駛行為的關鍵特徵；辨識為低風險之動作組合縱橫向陽性率分別為29.2%及6.9%，中高風險則上升至81.8%及47.7%，偽陽性大幅下降。整體而言，本研究所提之辨識流程對高風險駕駛行為具一定程度預測力，辨識結果可用以調整ADAS系統警示門檻，提高ADAS警示準確度。

Advanced Driver Assistance System (ADAS) is an essential vehicle equipment. An ADAS provides a real-time warning for drivers to pay attention to their own and surrounding vehicles. However, ADAS has been known for its false alarm issue, which, when the false alarm rate is high, would lead drivers to ignorance of ADAS warnings and thus raise the danger of driving. To resolve this issue, the study develops an ADAS-based risky driving behavior identification process; the identification results could be used to adjust the warning thresholds of ADAS alarms and enhance the accuracy of ADAS. The study first defines and labels dangerous events as the true values for ADAS warnings, followed by a binary logit analysis for selecting significant variables and a boosted regression tree analysis for examining possible nonlinear relationships between continuous explanatory variables with the onset of dangerous events. Accordingly, the continuous variables were discretized. The study applied the proposed framework to empirical data, which consisted of 576 longitudinal warnings and 425 lateral warnings. The results showed that the identified low-risk scenarios had a true alarm rate of 29.2% and 6.9%, respectively, for the longitudinal and lateral warnings, while the true alarm rate in the high-risk scenarios was 81.8% and 47.7%, respectively; the result validated the effectiveness of the proposed procedure. Overall, the proposed approach has a certain accuracy in predicting risky driving behaviors.