以卷積神經網路建構客運駕駛跟車距離之分類模式

Abstract

行車安全為政府所重視之公共運輸議題之一。若能了解客運駕駛行為之重要指標─駕駛跟車距離，業者便能夠有效地對駕駛員採取適當的行車管理措施。本研究之目的在以卷積神經網路建構一套跟車距離分類模型，將車載系統Mobileye所記錄之駕駛員跟車車距資料作為輸入資料，建立與訓練卷積神經網路分類模型，並藉由輸入之Mobileye車距資料來分類該駕駛員的跟車類型。本研究先將Mobileye所記錄之駕駛員跟車車距資料進行資料處理，並考量不同情境下之跟車距離，包括：天氣、車速及日夜三個因素，進一步設計七種不同的情境組合來篩選資料並產製車距矩陣圖，接著將車距矩陣圖輸入至卷積神經網路進行模型訓練及分類，利用訓練完成的模型來判斷所輸入之駕駛員車距圖片屬於何種跟車類型。實驗結果發現本研究所提出之分類模式在大多情境下皆有不錯的分類結果，其中以經圖片預處理之兩群分類準確率平均97%為最佳，顯示本研究之分類模型具備駕駛員跟車距離分類之能力，此模型可提供客運業者作為參考，藉由分類結果評估駕駛員的跟車距離類別，並針對不同類別駕駛員採行適當的管理措施。

Safety is one of the issues in public transportation that the government emphasizes. If bus companies are able to comprehend the most indicator of driving behavior of bus drivers—car-following distance, then companies can effectively adopt appropriate management measures for drivers. The purpose of this study is to develop classification models of bus drivers' car-following distance using convolutional neural networks (CNN). The headway data recorded by Mobileye systems were used as the input data to build and train the CNN classification models. The proposed classification models would facilitate determining the car-following types of bus drivers based on Mobileye headway data. We firstly processed the Mobile headway data and then considered different scenarios of car-following distance of drivers under three factors, including weather, speed and day-or-night. According to the three factors, we designed seven different scenarios and filtered the data by scenario. The filtered data were used to generate headway matrices for each driver. The matrices were the input data of the CNNs for training and classification. The trained models can be used to determine the car-following type of the input headway matrix of a driver. The experimental results show that the proposed CNN models have accurate prediction in most of the scenarios. Among them, the best classification results are the two groups that have been processed by the image preprocessor, and the accuracy rate is 97% on average. The results of this study can be used as a reference in classifying and determining the car-following distances of drivers. Bus companies can utilize the classification results to take actions on the driving behavior of drivers.