Inferencing transportation mode using unsupervised deep learning approach exploiting GPS point-level characteristics

Discovering the mode of transportation is a fundamental and challenging step in the various transportation analysis problems such as travel demand analysis, transport planning, and traffic management. Supervised learning approaches have frequently been used to detect the mode of transportation. In the supervised learning approach, extracted features from the GPS data and labeled mode of each trajectory are used to detect the mode of transportation. However, the sample size of trips for training is not enough due to the scarcity of labeled data. In real-time scenario, labeling the data is a tedious and expensive task. Furthermore, a huge amount of unlabeled data remains unutilized, which in turn degrades the quality of the solution. To overcome these drawbacks, an unsupervised learning method is proposed in this paper. In the proposed work, the point level characteristics such as speed, acceleration/deceleration, jerk, and bearing angle are fused with GPS coordinates to extract the joint probability densities using a type of generative deep learning model called masked autoregressive flow (MAF). Based on the probability density of each trajectory, K-means algorithm is applied to find the modes of transportation. The proposed model can handle the long sequence of trajectories having variable lengths. In this work, we compare our methodology with traditional machine learning algorithms and most related studies. The results are in favor of incorporating the proposed unsupervised approach in transportation mode detection.