Detecting travel modes from smartphone-based travel surveys with continuous hidden Markov models

In the last decades, studies on travel mode detection from location data have been increasing exponentially. However, these studies have struggled with three limitations: data collection-, feature selection-, and classification approach–related issues. Thus, we propose a novel framework to collect trajectory data and infer travel modes by making a great deal of effort. First, we conduct a travel survey with smartphones in Shanghai City, China. Furthermore, we use a prompted recall survey with surveyor intervention by telephones. In the survey, the surveyor asks respondents to validate the travel information automatically detected from trajectory data. Second, we use well-known sequential forward selection procedures to select the most reasonable combination of features. This set of features is expected to help achieve high classification accuracy with few features. Third, as a machine learning approach incorporating high resistance to noise in features, a continuous hidden Markov model is used to classify segments in dataset 1 that comprises Global Positioning System data alone. Consequently, 94.37% of segments are flagged correctly for the training dataset, while 93.47% are detected properly for the test dataset by making a comparison between detected travel modes and travel modes validated during the prompted recall survey. A higher accuracy (95.28%) is achieved in the test dataset on dataset 2 that consists of Global Positioning System, accelerometer, Global System for Mobile communication, and Wi-Fi data. The promising results obtained with this method provide a new perspective in understanding travel mode detection and other related issues in Global Positioning System travel surveys, including trip purpose detection.