Multimode trip information detection using personal trajectory data

Abstract

ABSTRACTHandheld global positioning system (GPS) devices can serve as a new tool to collect an individual's trip information with advantages of low cost, accurate data, and intensive spatial coverage. Various machine learning algorithms have been explored to detected trip train information in previous studies; however, few of them focused on the evaluation and comparison of the performance and applicability of different models. Meanwhile, according to previous studies, car and bus mode detection is a thorny issue due to their similar travel characteristics, and algorithms still need to be well explored and improved to solve this problem. In this article, an innovative method is proposed to detect trip information, including trip modes, mode-changing time and location, and other attributes, from personal trajectory data. The method is a two-step process. A machine learning algorith-based module (including artificial neural network, support vector machine, random forests, and Bayesian network) is firstly used to identify walk, bicycle, and motorized trip modes (bus or car); we thoroughly compared the performance of these four algorithms. Then a second module, using critical points on the GPS trajectories, is further developed to distinguish car and bus mode, incorporated with GIS map information. Field test results show that the proposed machine learning models can all be applied for walk, bicycle, and motorized mode detection with high detection rates exceeding 90%; however, the algorithms work relatively poorly for bus and car mode detection, with results mostly below 75%. The proposed two-step method can greatly improve bus and car mode detection accuracy by 14–30%. As a result, the average mode detection rates for all the four modes are above 90%. Compared with mode detection results by using only the machine learning algorithm, the proposed two-step method has much better performance in both accuracy and consistency.