請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/16899
標題: | 取消第三車道禁行機車安全評估-以深度學習模型分析 Deep Learning for Evaluating Deregulation of Motorcycle-Forbidden Policy on Third Lane |
作者: | Hsin-Hsuan Wu 吳欣璇 |
指導教授: | 許添本(Tien-Pen Hsu) |
關鍵字: | 深度學習,交通安全,事前事後分析,第三車道禁行機車,肇事預測, Motorcycle-Forbidden Policy,Before and After Analysis,Accident Prediction Model,Traffic Safety,Deep Learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 臺灣三車道以上路段普遍劃設「禁行機車」字樣,禁止機車行駛於內側車道,然而該政策為40年實施之政策,隨著時空推移道路環境已有所不同,因此台北市政府於民國98年起即開始採個案檢討方式,以例外開放原則陸續取消外側第三車道禁行機車之管制,然而針對該管制取消對於交通安全之影響尚無詳細之事前事後分析。 事前事後分析主要之主軸為將實際發生事故數與若未實施措施下之預期事故比較,因此本研究欲透過建立肇事事故數預測模型進行分析,可預測A1、A2與A3事故數後再換算EPDO,比較其肇事風險來判定改善或惡化。首先蒐集台北市已開放之64條路段的肇事資料,並利用深度學習中前饋式神經網路與卷積神經網路進行建模,同時也以多元負二項統計模型與隨機森林機器學習模型進行預測,並以均方誤差作為評估指標,比較其交叉驗證與測試集之預測誤差。交叉驗證之成績中,以隨機森林之表現最好,然而在測試集成績中,則以前饋式神經網路模型表現最好,而該兩者模型之預測誤差皆低於傳統多元負二項模型,顯示有其應用價值。兩預測模型的結果可用於過去路段之改善與惡化分析,以及評估其他未開放路段是否應取消管制。 在隨機森林與前饋式神經網路預測之結果中,380條子路段中共有284條預測結果相同,將這些子路段依改善惡化分類後觀察其特徵次數分布可發現惡化路段土地為商業使用者較多,而在大型車交通量與公車尖峰小時班次數方面,惡化路段之交通量皆較向數值高之方向集中,顯示較繁忙之路段若開放第三車道交通安全惡化之可能性較高。 In Taiwan, the road marking, “motorcycle-forbidden”, is common to be painted on the inner lane of three-lane roads. However, the motorcycle-forbidden policy has been implemented since 40 years ago, and the traffic conditions have also changed over the years. Consequently, the authorities concerned had deregulated the policy on the third lane case by case, allowing motorcyclists to ride on the inner lane; nevertheless, the traffic safety before and after analysis hasn’t been conducted thoroughly. The principle of the before and after analysis is to compare the real number of accidents and the expected one of accidents without treatment, and accordingly the study aims to develop an accident prediction model, which can predict the number of A1, A2 and A3 accidents. The evaluative criterion is to calculate as well as comparing the equivalent property damage only (EPDO) crash rate. The study collects the accident data of the 64 roads and develops four accident prediction model: the traditional statistic model, multivariate negative binomial model(MVNB); the classical machine learning model, random forest model(RF); the deep learning models, feed-forward neural network(FNN) and convolutional network(CNN). The performance of the models is evaluated by the mean squared error of cross validation and testing sets respectively. In cross validation, the score of RF is the best while FNN performs better than other models in the testing sets. What’s more, the prediction errors of these two models are lower than the one of MVNB. The result shows the RF and FNN have their application value. The prediction model can be applied to evaluate the traffic safety of the roads whether the restriction has been deregulated or not. From the prediction of RF and FNN, 284 out of 380 segments have the same conclusion of improvement and deterioration. Hence, the frequency distributions of the road features have been analyzed. For the road segments where the traffic safety worsens, the land use would more likely to be commercial use; the traffic volume of full-sized vehicles and peak-hour frequency of buses tend to concentrate on the larger value. The result indicates that after the deregulation of the motorcycles policy, the traffic safety might deteriorate on the heavy-traffic roads. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/16899 |
DOI: | 10.6342/NTU202002726 |
全文授權: | 未授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0908202018022500.pdf 目前未授權公開取用 | 4.71 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。