請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/49742
標題: | 基於LSTM神經網路之機車微觀車流模型 Microscopic Traffic Flow Modeling for Motorcycles using Long Short-Term Memory Neural Network |
作者: | Ting-You Lin 林亭佑 |
指導教授: | 許添本(Tien-Pen Hsu) |
關鍵字: | 微觀車流模擬,混合車流,機車,長短期記憶神經網路,深度學習, microscopic traffic simulation,mix traffic flow,motorcycles,long short-term memory neural network,deep learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 機車的機動性能與駕駛者感知判斷之間的關係相當複雜,且行為模式與汽車有相當的差異,導致典型以跟車模式與變換車道模式為基礎的微觀車流模擬軟體應用於機車組成比例大的道路環境表現不佳。此外,機車行為與駕駛因素有很大的關聯,卻難以量化成數學模式,許多國內外文獻藉由神經網路探索出非結構化的模式,並藉由神經網路中複雜的結構,獲取影響機車行為更細微的特徵因素,以有效的表達道路上車輛行為及駕駛因素的特徵。 根據文獻證實,深度學習演算法能有效地將車輛周圍複雜的車輛間互動與影響駕駛行為的因素特徵化,而遞迴神經網路或是其他改良式網路等同一類型的模型,適合用於預測如車流行為等時間序列上的空間移動行為。本研究將基於這些文獻,提出以長短期記憶神經網路為基礎的機車微觀車流模式。模式架構中以局部主體機車與各角度的周圍相鄰車輛變數組成多個全連接層,提取時間維度上各時點主體車輛與周圍車輛微妙的互動關係,接著以長短期記憶層擬合這些關係隨時間序列變化的資訊,最後以駕駛決策的加速度與行進方向作為模式輸出,以描述機車的微觀行為。 本研究的資料來源為高雄智慧機車計畫案的空拍車流資料,將車流影像轉換為軌跡資料後,得以用於訓練及驗證本研究提出的模式。研究範圍為路段中推進的行為模式。微觀驗證顯示本研究之模式有能力重現諸如超車等機車特有行為,而巨觀驗證足以證明模式的可行性。本研究除了預期提出穩健的車流模擬模式,亦提供針對在不同地點以神經網路為基礎的建模流程。 The relationship between the maneuverability of scooters and the driver's perception is quite complicated, and the behavior pattern is quite different from that of the car, which led to the poor performance of the typical microscopic traffic simulation software based on the car-following model and lane-changing model when applied to roads with a large proportion of scooters. In addition, scooter behavior is closely related to driver factors, but it is difficult to quantify it into a mathematical model. Many domestic and foreign literatures have used neural networks to explore unstructured patterns, and we can obtain more subtle characteristic factors that affect the behavior of the scooter to effectively express the characteristics of the vehicle behavior and driver factors on the road through the complex structure in the neural network. According to the literatures, the deep learning algorithm can effectively characterize the complex inter-vehicle interactions around the vehicle and the factors that affect driving behavior. The same type of model, such as recurrent neural networks or other improved networks, is suitable for predicting spatial movement behavior in time series such as driving behavior. Based on these literatures, this study will propose a microscopic traffic model based on long short-term memory neural networks. In the model architecture, multiple fully connected layers are composed of local subject scooter and neighboring vehicle variables at various angles to extract the subtle interaction relationship between subject scooter and surrounding vehicles at each time step in the time dimension. Then, the long short-term memory layers are used to fit the information of these relationships with the time series. Finally, the acceleration and driving direction of the driving decision are used as the model output to describe the microscopic behavior of the scooter. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/49742 |
DOI: | 10.6342/NTU202003033 |
全文授權: | 有償授權 |
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 | |
---|---|---|---|
U0001-1208202000054800.pdf 目前未授權公開取用 | 2.21 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。