Learning Traffic as Images: A Deep Convolutional Neural Network for Large-Scale Transportation Network Speed Prediction (1701.04245v4)

Abstract: This paper proposes a convolutional neural network (CNN)-based method that learns traffic as images and predicts large-scale, network-wide traffic speed with a high accuracy. Spatiotemporal traffic dynamics are converted to images describing the time and space relations of traffic flow via a two-dimensional time-space matrix. A CNN is applied to the image following two consecutive steps: abstract traffic feature extraction and network-wide traffic speed prediction. The effectiveness of the proposed method is evaluated by taking two real-world transportation networks, the second ring road and north-east transportation network in Beijing, as examples, and comparing the method with four prevailing algorithms, namely, ordinary least squares, k-nearest neighbors, artificial neural network, and random forest, and three deep learning architectures, namely, stacked autoencoder, recurrent neural network, and long-short-term memory network. The results show that the proposed method outperforms other algorithms by an average accuracy improvement of 42.91% within an acceptable execution time. The CNN can train the model in a reasonable time and, thus, is suitable for large-scale transportation networks.

• The paper introduces a CNN method that transforms spatiotemporal traffic data into image-like matrices, yielding a 42.91% accuracy improvement over traditional models.
• It details a convolutional architecture that extracts hierarchical features through multiple layers to address both short-term and long-term speed prediction.
• The study demonstrates scalable and robust real-time traffic management applications by validating the model on real-world networks in Beijing.

Learning Traffic as Images: A Deep Convolutional Neural Network for Large-Scale Transportation Network Speed Prediction

Overview

This paper introduces a convolutional neural network (CNN)-based methodology for predicting traffic speeds over large-scale transportation networks by representing spatiotemporal traffic data as images. The authors propose converting the dynamic traffic data into a two-dimensional time-space matrix and employing a CNN to extract spatiotemporal features and make predictions. The method was validated through empirical studies on two real-world transportation networks in Beijing. Performance comparisons were made against several traditional and deep learning approaches, demonstrating the superior accuracy and robustness of the proposed CNN-based method.

Methodology

Traffic data for the transportation networks were transformed into a time-space matrix, where the x-axis represents time intervals and the y-axis represents road segments. Each matrix element corresponds to the average traffic speed. This matrix was then treated as a grayscale image input to a CNN. The CNN architecture included multiple convolutional layers followed by pooling layers to extract hierarchical spatiotemporal traffic features, and a fully connected layer for prediction.

Key Results

1. Accuracy Gains: The proposed CNN outperformed all comparison models, including ordinary least squares (OLS), k-nearest neighbors (KNN), random forest (RF), artificial neural networks (ANN), stacked autoencoders (SAE), recurrent neural networks (RNN), and long short-term memory networks (LSTM NN). The CNN achieved an average accuracy improvement of 42.91%.
2. Efficiency: The CNN demonstrated reasonable training times, outperforming computationally intensive models such as RF without compromising accuracy.
3. Scalability: The CNN effectively handled large-scale network traffic predictions and exhibited robustness across various tasks involving short-term and long-term prediction horizons.

Implications

Practical Implications

1. Real-time Traffic Management: The CNN-based method can support real-time traffic prediction on large-scale road networks, enabling more efficient traffic management and route planning. By integrating spatiotemporal features, this approach can better anticipate traffic conditions and congestion patterns.
2. Resource Allocation: Traffic managers can leverage these predictions to allocate resources more effectively. For instance, by predicting areas likely to experience congestion, law enforcement and public transport can be preemptively deployed to mitigate potential delays.

Theoretical Implications

1. Spatiotemporal Feature Extraction: This paper underscores the importance of CNNs in extracting relevant spatiotemporal features for traffic predictions. The approach differentiates itself from traditional models that often treat traffic data as independent time series, thereby neglecting spatial correlations.
2. Future Research: This methodology sets a precedent for further exploration into hybrid models combining CNNs with RNN and LSTM architectures to enhance prediction capabilities. Specifically, combining the feature extraction strengths of CNNs with the temporal sequencing capabilities of RNNs and LSTMs may yield even higher accuracies and more robust models.
3. Adaptation to Other Domains: The approach of converting spatiotemporal data into image-like representations can be adapted to other domains requiring spatiotemporal analyses, such as weather forecasting, stock market predictions, and healthcare analytics.

Conclusion

This paper advances the field of traffic prediction by showcasing the efficacy of CNNs in handling large-scale, spatiotemporal traffic data. By transforming traffic data into images and leveraging deep learning, the authors have presented a method capable of producing highly accurate and efficient traffic predictions. The promising results encourage further exploration into deep learning models that can capture complex data relationships, potentially leading to improved modeling and forecasting in various application domains.