Bayes Neural Network with a Novel Pictorial Feature for Transportation Mode Recognition Based on GPS Trajectories

Recognizing transportation mode from raw trajectory is a crucial task in transportation, which benefits travel pattern understanding and traffic management. However, difficulties such as small, unbalanced datasets and extracting efficient features from raw trajectory data persist. This paper proposes a novel image-based feature to represent trajectory, since we find that various transportation modes are easily distinguished in image representation, preserving original trajectory information. The feature's low rank nature also improves overfitting problems. Additionally, due to heavily unbalanced datasets, traditional classifiers, such as support vector machines, are difficult to perform. To mitigate this, we introduce a Bayes neural network, learning a distribution for each parameter instead of a single value. Results show that our novel image-based feature improves the overfitting problem, and the combination of the image-based feature and Bayes neural network achieves competitive performance on all modes, especially obtaining performance improvements on subways and trains by 13.4% and 42.8%, respectively.