Travel Mode Detection Based on GPS Data: Confusion Impacts and Accuracy Improvement Strategies

The smartphone-based GPS data collection has become an innovative means of household travel surveys. Still, more is needed to replace the conventional interview-based surveys for the limited performance in detecting travel mode and trip purpose. With much attention drawn to the application of machine learning models, previous studies ignored the distinction among ambiguous transports, which, however, is the key to improving the accuracy of travel mode detection under certain circumstances. This paper first compared the performance between three models in detecting five transport modes, including walking, bicycle, bus, car, and most importantly, e-bike, which is recently emerging and rapidly developed as a dominant transport for short-distance trips in China and was rarely considered before. Then, the confusion among modes is stressed and figured out with model selection and feature engineering strategies. Results show that the confusion between bicycle and e-bike and between bus and car is difficult to be distinguished due to the similar mobile characteristics. But the accuracy is significantly improved to 89.2% with additional features added. The appropriate model and features are vital to enhancing travel mode detection, which would, in turn, promote the development of travel survey techniques.