以深度學習長短時記憶神經網路演算法建立混合車道之機車行為模式

A Study of Using Deep Learning Method(Long Short-Term Memory) to Build Motorcycle Moving Behavior on Urban Mixed Lane

吳建霆(Chien-Ting Wu)

指導教授 ： 范俊海

淡江大學/商管學院/運輸管理學系運輸科學碩士班/碩士(2018年)

https://doi.org/10.6846/TKU.2018.00987

摘要

機車為台灣最主要的交通工具之一，特色為行進時不會遵行固定路線，經常會於同一車道內併行或連續超車，在過去針對機車行為研究中，試圖利用單一數學模型解釋機車推進模式及連續超車的各種行為。但本研究認為變化萬千的機車行為模式，若只利用單一模型解釋可能會造成模擬上的困難，且無法體現台灣車流混雜的道路真實情況，故本研究針對台灣機車的特殊行車狀態，於一般市區道路汽機車混合車道中的機車行為模式為研究主體。
本研究利用深度學習運算演算法，建立機車行為模式預測模型，將機車的前進方向及影響機車推進的各項影響參數，建立其輸入與輸出變數的對應關係，並以攝影座標轉換法擷取道路機車座標，以求得更準確之研究數據，為了讓預測結果接近於數值型的座標預測，利用深度學習特殊的運算結構，結合座標及格位概念預測機車的推進行為，研究中將機車的推進行為，依座標詳細計算25個推進模式，配合25個影響行為參數，在連續的時間序列下建立關係矩陣，設計深度學習的運算結構規則。研究結果，透過模式建立成功將機車行為分成25個行進模式，並利用資料後段之25%與前段75%進行模式驗證，模式預測程度可達84%，證明模式預測機車行為確實有可行性，而後續研究可根據本研究之研究成果，蒐集完整的車流資訊執行模擬訓練，以建立完整的車流特性資料庫。

關鍵字

機車行為模式 ； 深度學習 ； 預測 ； 座標轉換 ； 長短時記憶神經網路

並列摘要

Motorcycle is one of the Taiwan main transportation vehicles, they cannot follow the specific route, and always frequently parallel or overtake in the identical traffic lane. In the past research about motorcycle moving behavior that tring to use a single mathematical model to explain the Motorcycle moving behavior models and the moving behaviors of continuous overtaking. However, this paper considered that it’s difficult to explain and simulate the changeable motorcycles behavior pattern by a single model, and it's unable to reflect the true situation of the mixed urban road in Taiwan. Therefore, this paper is aimed at special driving state of the Motorcycles in Taiwan and behavior pattern in the mixed road of the general urban road is the main research subject.
This paper builds a predictive model of motorcycles moving behavior patterns, using the arithmetic structure with deep learning method. Establish the corresponding relationship between input and output variables of the motorcycles' forward direction and various influence parameters affecting the behavior. To obtain the complete traffic coordinates, using the coordinates of motorcycles with the camera coordinate conversion method, in order to make the prediction result approach the numerical coordinate prediction, using the special arithmetic structure of deep learning integrate coordinate and lattice concept. In the study, the motorcycles moving behavior is calculated in detail according to the coordinates of 25 propulsion modes ,with 25 influencing behavior parameters, establish a relational matrix in a continuous time series and design the operational structure rules for deep learning .The model result shows that through the establishment of the mode, the motorcycles moving behavior was successfully divided into 25 travel modes, finally to perform pattern verification, divide 25% section of the data ,and verify another 75% . The mode prediction degree can reach 84%, which proves that the model predicts the motorcycles moving behavior is indeed feasible. Subsequent research can collect complete traffic information and perform simulation training based on the results of this research to establish a complete vehicle flow characteristic database.

並列關鍵字

Driving Behavior of Motorcycles ； Deep learning ； Prediction ； Coordinate ； Long short-term memory Network

參考文獻

參考文獻

1. Ahmed, K., Ben, A.M., Koutsopoulos, H.N., Mishalani, R.G. (1996). Models for freeway lane changing and gap acceptance behavior. Proceedings of 13th International Symposium on Transportation and Trajfic Theory, Lyon France.

2. Ahmed, K. (1999). Modeling drivers’ acceleration and lane changing behavior. Unpublished doctoral dissertation, Massachusetts Institute of Technology.

3. Abdelhadi, A., Guy, P. (2007). A Long Short-Term Memory Recurrent Neural Network Framework for Network Traffic Matrix Prediction. LIP6 / UPMC. Paris, France.

4. Carlos, F.D. (1981). Estimation of gap acceptance parameters within and across the population from direct roadside observation, Transportation Research Part B vol. 15. pp.1-15.