Travelling modes recognition is essential to traffic and infrastructure management. Nevertheless, adverse factors such as lacking of training data, unbalanced trip mode data, and undetermined input features, hinder the process of obtaining a precise and robust travelling modes recognition. In this paper, we develop a novel method using Bayes neural network (BNN) with Bayes by Backprop algorithm to solve the problem. First, the trajectories are segmented into sub-trajectories in terms of the various transportation mode. Then, a comprehensive feature called velocity statistics histogram feature (VSHF) would be built as the input of our BNN. Each parameter of BNN is a distribution rather than a fixed constant, so it can escape from involving the dilemma of overfitting. Besides, it has a strong capability to predict the unseen data, and further improving the prediction performance. The experiments showed that BNN can significantly improve the performance of traditional neural networks in travelling mode recognition. Our method can also be applied to other tasks requiring accuracy improvement or robustness increment.

CICTP