Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tel2Veh: Fusion of Telecom Data and Vehicle Flow to Predict Camera-Free Traffic via a Spatio-Temporal Framework

ChungYi Lin, Shen-Lung Tung, Hung-Ting Su, Winston H. Hsu

TL;DR

This work proposes a two-stage framework that first independently extracts features from multimodal data and then integrates them using a graph neural network (GNN)-based fusion to generate predictions of vehicle flow in camera-free areas and pioneer the fusion of telecom and vision-based data.

Abstract

Vehicle flow, a crucial indicator for transportation, is often limited by detector coverage. With the advent of extensive mobile network coverage, we can leverage mobile user activities, or cellular traffic, on roadways as a proxy for vehicle flow. However, as counts of cellular traffic may not directly align with vehicle flow due to data from various user types, we present a new task: predicting vehicle flow in camera-free areas using cellular traffic. To uncover correlations within multi-source data, we deployed cameras on selected roadways to establish the Tel2Veh dataset, consisting of extensive cellular traffic and sparse vehicle flows. Addressing this challenge, we propose a framework that independently extracts features and integrates them with a graph neural network (GNN)-based fusion to discern disparities, thereby enabling the prediction of unseen vehicle flows using cellular traffic. This work advances the use of telecom data in transportation and pioneers the fusion of telecom and vision-based data, offering solutions for traffic management.

Tel2Veh: Fusion of Telecom Data and Vehicle Flow to Predict Camera-Free Traffic via a Spatio-Temporal Framework

TL;DR

This work proposes a two-stage framework that first independently extracts features from multimodal data and then integrates them using a graph neural network (GNN)-based fusion to generate predictions of vehicle flow in camera-free areas and pioneer the fusion of telecom and vision-based data.

Abstract

Vehicle flow, a crucial indicator for transportation, is often limited by detector coverage. With the advent of extensive mobile network coverage, we can leverage mobile user activities, or cellular traffic, on roadways as a proxy for vehicle flow. However, as counts of cellular traffic may not directly align with vehicle flow due to data from various user types, we present a new task: predicting vehicle flow in camera-free areas using cellular traffic. To uncover correlations within multi-source data, we deployed cameras on selected roadways to establish the Tel2Veh dataset, consisting of extensive cellular traffic and sparse vehicle flows. Addressing this challenge, we propose a framework that independently extracts features and integrates them with a graph neural network (GNN)-based fusion to discern disparities, thereby enabling the prediction of unseen vehicle flows using cellular traffic. This work advances the use of telecom data in transportation and pioneers the fusion of telecom and vision-based data, offering solutions for traffic management.
Paper Structure (10 sections, 2 equations, 5 figures, 4 tables)

This paper contains 10 sections, 2 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Tel2Veh Dataset
  3. Definitions
  4. Data Collection and Processing
  5. Data Analysis
  6. Spatio-Temporal Fusion Framework
  7. Experiments
  8. Experimental Settings and Baselines
  9. Experimental Results
  10. Conclusion

Figures (5)

  • Figure 1: Overview of the task and dataset. (a) We propose a task that combines cellular traffic (i.e., mobile user network activities) with sparse vehicle flows to forecast vehicle flow in camera-free areas, supported by the Tel2Veh dataset which contains cellular traffic on roads and camera-detected vehicle flows. (b) Our framework, after training, has proven its capability for accurate prediction in camera-free areas.
  • Figure 2: Overview of GCT and vehicle flows in Hsinchu City. (a) Spatial distribution of 49 GCT and 9 camera-detected vehicle flows on various road segments. (b) GCT flows show unique temporal dynamics related to functional areas like the Science Park, and Commercial Area. (c) Comparison of GCT flows on Road Segment 5 and vehicle flow from Camera 1 at the same location, highlighting trend similarities.
  • Figure 3: Correlation between GCT and vehicle flows. (a) Daily Pearson correlation coefficients at matching locations generally indicate moderate to high correlations, implying pattern alignment, while low correlations suggest data anomalies. (b) An example of low correlation, such as the missing vehicle flow for segment 22 with Cam 9 after 14:00, suggests potential detection errors or device malfunctions.
  • Figure 4: Overview of the two-stage fusion framework. Stage 1 uses two pre-trained STGNNs for feature extraction from GCT and vehicle flows. Stage 2 integrates the extracted features via the Fusion Model, feeding them into the third STGNN for prediction. Only Stage 2 is trained, optimizing the Loss Function to align predictions with vehicle flows.
  • Figure 5: Daily 15-minute forecasts on Road ID 37 for 9/26. Ground truth (i.e., actual vehicle flow from Cam7) was excluded from training to serve as validation. The closer alignment of solid lines with the ground truth confirms the accuracy enhancement by our framework.