Deep Learning for Trajectory-Based Transportation Mode Identification
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Understanding users' mobility patterns and associated transportation modes is essential for intelligent transportation management and infrastructure design. Through the ubiquity of Global Positioning System (GPS) sensors in modern smartphones and vehicles, rich spatiotemporal trajectories can be readily captured for use in intelligent transportation applications. Key among the latter is transportation mode identification, or how to infer travel modes within GPS trajectories. Although studied extensively, its real-world applicability remains limited due to several challenges.
GPS trajectories are often incomplete due to signal lapses, thereby complicating subsequent analysis. Since learning from raw GPS data restricts model generalization to the regions best covered in the training set, this thesis sets an alternative imputation target: approximate missing GPS points by learning to impute relative magnitude and angle of displacement features. The proposed Uncertainty-aware Imputation Generative Adversarial Network (UI-GAN) leverages a Bayesian generator to capture imputation uncertainty and a window-level discriminator for localized sequence structure penalization. UI-GAN produces high-fidelity GPS points and outperforms established imputation baselines.
A single GPS trajectory may encompass multiple transportation modes. Existing trajectory segmentation approaches often exhibit poor scalability and require extensive feature engineering or transportation domain knowledge. As such, this thesis reframes trajectory segmentation as timestep-level transportation mode identification. Concretely, it proposes a shuffling-based data augmentation scheme and a majority-vote post-processing step to effectively train a convolutional neural network for timestep-level classification and refine the extracted segments. The proposed segmentation model is nearly twice as accurate as the best performing baseline in detecting transportation mode changes.
In reality, GPS trajectories are neither automatically annotated nor segmented by transportation mode. In addition, predictive uncertainty tied to model parameters or noise in GPS readings is typically unaccounted for. Therefore, this thesis proposes an unsupervised channel-calibrated Bayesian Temporal Convolutional Network (BTCN) trained to maximize the mutual information between neighboring feature map patches. By approximating variational inference, BTCN can both classify each input timestep and estimate its predictive uncertainty. BTCN significantly outperforms established trajectory segmentation baselines without using any labels.
Finally, this thesis proposes an unsupervised deep learning approach to transportation mode identification. First, a clustering layer maintaining cluster centroids as trainable weights is attached to the embedding layer of a convolutional autoencoder. The composite model is then trained by optimizing a weighted sum of reconstruction and clustering losses to encourage learning clustering-friendly representations. By further incorporating segment-level features, the proposed model outperforms traditional clustering and state-of-the-art semi-supervised methods without using any labels.
Please use this identifier to cite or link to this item: