以機器學習演算法評估客運駕駛風險之研究

Research on the evaluation of bus drivers’ driving risk based on machine learning algorithm

王允(Wang, Yun)

指導教授 ： 王晉元

國立交通大學/管理學院/運輸與物流管理學系/碩士(2020年)

摘要

本研究透過大客車配載之先進駕駛輔助系統(ADAS)記錄之車機資料，從中提取駕駛特徵，以機器學習演算法建構駕駛風險評估模型來預測其風險層級。機器學習演算法包含以非監督式學習之分群法根據風險指標來標記風險等級，在使用監督式學習之分類法連結駕駛行為指標建立風險分類模型，並以recursive features elimination演算法依照特徵的重要性找出關鍵特徵。另外針對過去研究使用2階機器學習架構和區分駕駛特徵之兩個議題進行探討，建立3種分類模型來測試其必要性和準確度。本研究之測試範例為行駛國道由新竹至台北的國內客運公司提供Mobileye車距資料，透過k-means分群法將駕駛分別出風險等級，並找出關鍵駕駛行為特徵。結果顯示3種模型之正確率皆高於90%且3種模型的準確度差異不大。意味著駕駛風險之評定取決於風險標籤所使用之駕駛特徵，將駕駛風險分群後再進行第2階段使用分類法建立風險分類模型並沒有其必要性；相同的將特徵分為風險行為兩類型對模型評估的準確度並無差異。因此相對其他模型，1階段混合特徵評估模型(Model-3)僅進行一次機器學習，較能避免球員兼裁判之情況。若考量日後資料蒐集和處理以及模型再訓練之時間，一階段混合特徵預測模型為較佳的選擇。

關鍵字

機器學習 ； 駕駛風險 ； 駕駛行為 ； 先進駕駛輔助系統

並列摘要

This study extracts driving features from the vehicle data by the advanced driver assistance system (ADAS) on the bus ,and uses machine learning algorithms to construct a driving risk assessment model. The machine learning algorithms include unsupervised learning (clustering) to label risk levels by risk features, and the supervised learning (classification) that use driving behavior features to establish risk classification models. Also, using recursive feature elimination algorithm to identify key features. Moreover, the two topics of using the second-order machine learning architecture and distinguishing driving characteristics in the past research were discussed, and three risk assessment models were established to test necessity and model accuracy. As a case study, using Mobileye data from domestic highway bus carrier, which operates from Hsinchu to Taipei. The k-means clustering method is used to cluster the driving risk levels and then find out the key driving behavior features by RFE. The results show that the accuracy of the three models are higher than 90% and not much different. It means that the assessment of driving risk depends on the driving features used by the risk label. So, it is unnecessary to carry out the risk classification model after risk label, divided features into two types neither. Therefore, among three models, the one-order mixed feature model (Model-3) only performs machine learning once, which can better avoid the situation of being players and referees in the same time. Considering the time for data collection and processing and the time of model retraining in the future, the one-order mixed feature model is the better choice.

並列關鍵字

Machine learning ； Driving risk ； Driving behavior ； ADAS

參考文獻

[1] Bell, J. L., Taylor, M. A., Chen, G. X., Kirk, R. D., & Leatherman, E. R. (2017). Evaluation of an in-vehicle monitoring system (IVMS) to reduce risky driving behaviors in commercial drivers: Comparison of in-cab warning lights and supervisory coaching with videos of driving behavior. Journal of safety research, 60, 125-136.

[2] Cheng, Z., Wang, W., Lu, J., & Xing, X. (2018). Classifying the traffic state of urban expressways: A machine-learning approach. Transportation Research Part A: Policy and Practice.

[3] Feng, S., Li, Z., Ci, Y., & Zhang, G. (2016). Risk factors affecting fatal busaccident severity: Their impact on different types of bus drivers. Accident Analysis & Prevention, 86, 29-39.

[4] Lourens, P. F., Vissers, J. A., & Jessurun, M. (1999). Annual mileage, driving violations, and accident involvement in relation to drivers’ sex, age, and level of education. Accident Analysis & Prevention, 31(5), 593-597.

[5] Ma, Y., Tang, K., Chen, S., Khattak, A. J., & Pan, Y. (2020). On-line aggressive driving identification based on in-vehicle kinematic parameters under naturalistic driving conditions. Transportation Research Part C: Emerging Technologies, 114, 554-571.