The 20th and 21st joint COTA International Conference of Transportation Professionals (CICTP 2020-21)

Recognizing the transportation mode from a raw trajectory is a fundamental but crucial task in transportation field, which can benefit to travelling pattern understanding and traffic management significantly. However, some difficulties, such as how to deal with unbalanced dataset or small dataset, and how to extract an efficient feature from raw trajectory data, make it remain open. To bridge this gap, in this paper, we propose a novel image-based feature to represent the trajectory, since we simply find that various transportation modes can be distinguished easily in an image representation. Also, image-based feature can preserve as much as possible information concerning the original trajectory, and low rank nature of the feature can improve the overfitting as well. In addition to that, due to heavily unbalanced dataset, traditional classifier, such as support vector machine, is extremely difficult to perform well on all modes. To end that, we introduce Bayes neural network, learning a distribution for each parameter instead of a single value, to mitigate this problem effectively. The results of experiments show that our novel image-based feature can improve the overfitting significantly, and the combination of the image-based feature and Bayes neural network can achieve competitive performance on all modes, especially obtaining performance improvements on subway and train by 13.4% and 42.8%, respectively.

CICTP