在各式的道路環境下，十字路口是其中非常容易引發交通事故的地點。為了降低路口事故的發生率並增進道路安全，本篇論文提出一車載路徑預測系統。當車載行駛接近至十字路口，該系統能及時預測車輛的未來路徑，以即時警示駕駛事故發生的可能性。本篇論文採用簡易且價格低廉的GPS感測器採收車載行駛至十字路口的軌跡資料，並將軌跡資料中的方位角轉換為偏轉角，再整合車速與偏轉角來產生訓練資料。與此同時，我們實作了Borderline-SMOTE來增加少數類別的占比，解決類別不平衡問題。在本篇的演算法中，我們使用隨機森林與自適應增強兩種集成學習演算法來訓練模型，並且實作交叉驗證來降低過適問題的發生機率。最後的實驗結果指出，隨機森林的效能優於自適應增強，而自適應增強的效能優於基礎分類器(決策樹)。;Intersections have long been known as the place where major traffic accidents most likely to happen. To reduce the number of accidents, this thesis proposes a vehicle path prediction system to predict the future direction when a vehicle is about to cross an intersection. The simple and inexpensive sensor, such as GPS sensor, is used to collect the dataset of vehicle trajectories at intersections. The trend of vehicle movements are derived from the heading in the trajectories, which is then combined with the vehicle speed to generate training data. To deal with the class imbalance problem, Borderline-SMOTE is implemented to increase the proportion of the minority class. In our path prediction algorithm, two ensemble learning algorithms, i.e., Random Forests and AdaBoost, are adopted for model training . The cross validation technique is used to avoid the issue of overfitting. The experiment results indicate that the Random Forest algorithm exhibits the best performance, and the Adaboost algorithm performs better than the base learner (i.e., Decision Tree).
