[Paper Review] A Capsule Network for Traffic Speed Prediction in Complex Road Networks

Youngjoo Kim, Peng Wang+2 more|arXiv (Cornell University)|Jul 23, 2018

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper proposes a capsule network-based deep learning model for traffic speed prediction in complex road networks, replacing traditional CNN max pooling with dynamic routing to preserve spatial-temporal relationships in spatio-temporal traffic data. Evaluated on real-world data from Santander, Spain, the method reduces root mean squared error by 13.1% compared to state-of-the-art CNN-based approaches.

ABSTRACT

This paper proposes a deep learning approach for traffic flow prediction in complex road networks. Traffic flow data from induction loop sensors are essentially a time series, which is also spatially related to traffic in different road segments. The spatio-temporal traffic data can be converted into an image where the traffic data are expressed in a 3D space with respect to space and time axes. Although convolutional neural networks (CNNs) have been showing surprising performance in understanding images, they have a major drawback. In the max pooling operation, CNNs are losing important information by locally taking the highest activation values. The inter-relationship in traffic data measured by sparsely located sensors in different time intervals should not be neglected in order to obtain accurate predictions. Thus, we propose a neural network with capsules that replaces max pooling by dynamic routing. This is the first approach that employs the capsule network on a time series forecasting problem, to our best knowledge. Moreover, an experiment on real traffic speed data measured in the Santander city of Spain demonstrates the proposed method outperforms the state-of-the-art method based on a CNN by 13.1% in terms of root mean squared error.

Motivation & Objective

• To address the limitation of convolutional neural networks (CNNs) in preserving inter-relationships in traffic data due to max pooling operations.
• To explore the application of capsule networks—originally designed for computer vision—on time series forecasting in transportation systems.
• To model spatio-temporal dependencies in traffic speed data from sparsely distributed induction loop sensors more effectively.
• To improve prediction accuracy in complex road networks by capturing hierarchical relationships in traffic patterns.
• To demonstrate the superiority of capsule networks over CNNs in traffic speed forecasting using real-world data.

Proposed method

• Transforms spatio-temporal traffic speed data into a 3D representation with spatial and temporal dimensions, treating it as an image for deep learning processing.
• Replaces traditional max pooling in CNNs with dynamic routing in capsule networks to preserve spatial relationships and feature hierarchies.
• Uses capsules to encode pose parameters (e.g., position, scale) of traffic patterns, enabling attention to spatial dependencies across road segments.
• Employs iterative dynamic routing to determine the strength of connections between capsules, allowing the network to learn meaningful relationships between traffic features.
• Trains the capsule network end-to-end on real traffic speed data to predict future speeds at various road segments.
• Applies the model to a real-world dataset from Santander, Spain, to evaluate performance against CNN baselines.

Experimental results

Research questions

• RQ1 Can capsule networks outperform convolutional neural networks in traffic speed prediction by preserving spatial-temporal relationships?
• RQ2 How effective is dynamic routing in modeling inter-relationships among traffic data from sparsely located sensors?
• RQ3 To what extent does replacing max pooling with capsule-based routing improve prediction accuracy in complex road networks?
• RQ4 Is the capsule network architecture suitable for time series forecasting tasks in transportation systems?
• RQ5 How does the proposed method compare to state-of-the-art CNN-based approaches in terms of root mean squared error on real traffic data?

Key findings

• The proposed capsule network achieves a 13.1% lower root mean squared error (RMSE) than the state-of-the-art CNN-based method on real traffic speed data from Santander, Spain.
• The model demonstrates superior performance in capturing long-range dependencies and spatial relationships in traffic patterns due to dynamic routing.
• The use of capsules effectively preserves feature hierarchies and spatial relationships that are lost in max pooling operations of standard CNNs.
• The method is the first known application of capsule networks to time series forecasting, particularly in traffic speed prediction.
• The results confirm that dynamic routing enhances feature representation learning in spatio-temporal traffic data, leading to more accurate predictions.
• The model maintains high performance even with sparse sensor data, indicating robustness in real-world deployment scenarios.