| 摘要: | 行動手機資料已被廣泛運用在旅運行為的範疇中，其中尤以運輸規劃四步驟：旅次產生、旅次分佈、運具判別、交通量指派最為常見。相較於旅運需求預測其他三種步驟，運具判別的議題現有文獻較少探討，故本研究特別針對運具判別進行探討。目前行動電話資料的點位產生方式係由行動電話主動傳遞訊號至基地台，並利用三角運算推估經緯度而得，但受限於現有的技術常存在兩個缺點：(1)時間與空間上的不確定性; (2)飄移現象。其中又可細分出飄移點、回跳點、資料點未位於道路路網上等問題，本研究將針對此三種偏差進行修正，其中本研究提出了一種新穎的方式以解決回跳點問題；而資料點匹配至道路之處理，則利用隱馬可夫模型進行圖形比對。為此，本研究採用兩種監督式分類方法(支持向量機與深度神經網路)，探討在四種運具特性情境(依據旅行時間、起迄對的開始時間、起迄對間的旅行速度、最大速度以及平均速度等五種特性組合而成)、四種資料集組合(兩對不同方向之起迄資料、以及根據東經121.206切割出兩種不同都市化程度之區域)、三種訓練驗證資料之比例、和資料是否進行圖形比對，進行系統性的測試，比較運具判別(公共汽車或是汽車)準確率上的表現。
研究結果顯示，SVM方法在(1)包含五個運具特性的情境；(2)從中央大學出發之起迄對；(3)60/40比例之訓練驗證資料；(4) 有或無考慮圖形比對，所進行的訓練結果，在運具判斷的表現上較其他組合優異，其準確率皆高達99.89%。但使用五個運具特性組合(情境一)的進行訓練存在實務上的限制：必須同時記錄實際的蒐集情形以獲得該起迄對的旅行時間。在資源有限且不特別萃取特定方向的前提下並不容易做到，因此可以採用不考慮旅行時間(情境三)的方式進行訓練，其訓練之準確率分別為80.38%(有圖形比對)、81.36%(無圖形比對)。前者加入圖形比對之效果雖稍低於後者，還是具備可以接受的準確率。行動電話資料經過圖形比對後會提升資料的正確性，但也同時造成運具間的速度特性相似度變低(因部分中速之距離經圖形比對後拉長了，兩運具的平均速度與速度之眾數皆分散至不同速度區間而導致兩者相差度降低，參見表5與圖20)，而增加特徵萃取難度。但此現象隨電信通訊技術的進步，可以預期利用行動電話作為運具判別之資料其正確性與資料點分佈密集程度將能有所提升，從而增加判別之準確性。
;Mobile phone data have been extensively explored for travelers’ behavior in the area of transportation planning, especially for extracting input data to trip generation, O-D estimation, mode choice and traffic assignment. In this research, we particularly address the inference of transportation modes using mobile phone data, because, as compared with the other three steps in the traditional travel demand forecasting procedure, research on extracting input data for mode choice using mobile phone data is relatively inadequate and thus the research gap need to be fulfilled. The obstacle that prevents the usage of mobile phone data from inferring the transportation mode is mainly due the two types of drawbacks, temporal and spatial uncertainties, and oscillation phenomenon. As a result, inaccurate locations of data points such as missing data, drifting data points, jump-back data points, and off-road data points appear and hence prevent the mobile phone data from being practically used. Among all the undesired phenomenon, this research will stress on the problem of jump-back and off-road data points, along with other influential factors such as level of urbanization (divided by an artificial separating line) which makes the layout and density of cell towers as well as traffic flow pattern totally different. A novel method was proposed to deal with jump-back data points whereas a map-matching method using hidden Markov model was adopted to treat off-road data points. For comparison purpose we also adopt two classification methods, i.e., support vector machine and deep neural network, to investigate how (1) the four scenarios formed based on the five modal features (including travel time, starting time of trace, traversal speed between traces, maximum speed, and average speed), (2) four combinations formed by the two directional O-D pairs and two subareas separated by longitude 121.206, (3) three proportion settings of training/testing data, and (4) map-matching techniques (with or without) would affect the accuracy rates associated with two training methods, i.e., SVM and DNN methods, in inferring transportation modes (i.e., either bus or vehicle).
The results show that the refined mobile phone data with (1) the first scenario consisting of five modal features, (2) the directional O-D data beginning from National Central University, (3) the proportion setting of (60, 40), (4) SVM, and (5) whether or not considering map-matching, resulted in the best performance with accuracy index of 99.89%. However, the feature ‘‘travel time’’ in the first scenario can only be collected by the investigator in the field survey. Therefore, the approach takes instead the third scenario (without considering travel time) for later analysis and can still achieve accuracy rate of, respectively, 80.38% for incorporating map-matching treatment, and 81.36% for without incorporating map-matching treatment. The performance of the former is a little inferior to the latter, however, it is still acceptable because the quality of the refined data points is still not justified which may affect the performance of later analysis. Theoretically the refined data after map-matching would improve the accuracy of data location, however, the speed distributions between the two types of transportation modes will thus become more similar, due to part of shorter distance paths corresponding to the statistical mode of speed distribution before map-matching were projected onto longer ones. Consequently, the extraction of useful modal features for inferring transportation modes becomes more difficult. Nevertheless, with the expected advances in mobile phone technology, the accuracy of locations obtained from and the frequencies associated with mobile phone data will improve and can of course get better performance in terms of accuracy rates. |
