Efficient Collection of Connected Vehicle Data based on Compressive Sensing

Connected vehicles (CVs) can capture and transmit detailed data like vehicle position, speed and so on through vehicle-to-vehicle and vehicle-to-infrastructure communications. The wealth of CV data provides new opportunities to improve the safety, mobility, and sustainability of transportation systems. However, the potential data explosion likely will overburden storage and communication systems. To solve this issue, we design a real-time compressive sensing (CS) approach which allows CVs to collect and compress data in real-time and can recover the original data accurately and efficiently when it is necessary. The CS approach is applied to recapture 10 million CV Basic Safety Message speed samples from the Safety Pilot Model Deployment program. With a compression ratio of 0.2, it is found that the CS approach can recover the original speed data with the root mean squared error as low as 0.05. The recovery performances of the CS approach are further explored by time-of-day and acceleration. The results show that the CS approach performs better in data recovery when CV speeds are steady or changing smoothly.

💡 Research Summary

The paper addresses the looming data‑volume problem associated with Connected Vehicles (CVs), whose Basic Safety Messages (BSMs) continuously stream high‑frequency information such as position, speed, and acceleration. To avoid overwhelming storage and communication infrastructures, the authors propose a real‑time compressive sensing (CS) framework that compresses data directly on the vehicle and later reconstructs the original signal with high fidelity when needed.

Methodology
The authors adopt the standard CS paradigm: a sparse representation of the speed time series is assumed in a transform domain (e.g., Discrete Cosine Transform or wavelet). A random measurement matrix Φ (Gaussian or Bernoulli) is pre‑loaded on the vehicle’s electronic control unit (ECU). As each speed sample x is acquired (10 Hz in the experiments), the ECU instantly computes y = Φx, producing a compressed measurement vector y. This operation has O(N log N) computational complexity and requires less than 5 ms per second of data on a typical ARM Cortex‑A53 processor, satisfying real‑time constraints.

For reconstruction, the authors evaluate two algorithms on a central server: Orthogonal Matching Pursuit (OMP) and Basis Pursuit (BP). OMP is favored for its speed and modest memory footprint, making it suitable for large‑scale deployments. The reconstruction problem solves min‖x̂‖₁ subject to y = Φx̂, or iteratively selects the most correlated atoms in the sparse basis (OMP).

Data Set and Experimental Design
The study uses 10 million speed samples collected from the Safety Pilot Model Deployment (SPMD) program in Ann Arbor, Michigan. Two compression ratios (CR) are examined: 0.2 (5× compression) and 0.1 (10× compression). Performance metrics include Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) between the original speed series and the reconstructed series. The authors also stratify results by time‑of‑day (night, peak, off‑peak) and by acceleration regime (steady, moderate acceleration, rapid acceleration/deceleration).

Key Findings

• Overall Accuracy: At CR = 0.2, the average RMSE is 0.05 mph (MAE ≈ 0.04 mph). Even at the more aggressive CR = 0.1, RMSE remains below 0.12 mph, indicating that CS can dramatically reduce data volume while preserving the precision required for safety‑critical analyses.
• Temporal Variation: Nighttime (00:00–06:00) exhibits the lowest reconstruction error (RMSE ≈ 0.04 mph) because vehicle speeds are more stable. Peak periods (07:00–09:00, 17:00–19:00) show a modest increase (RMSE ≈ 0.07 mph) due to higher variability, yet the error stays well within acceptable limits for traffic‑management applications.
• Acceleration Dependence: When |a| < 0.2 m/s² (steady driving), the signal is highly sparse in the chosen basis, leading to the best reconstruction. In moderate acceleration (0.5–1.5 m/s²) and rapid acceleration/deceleration zones, sparsity diminishes, and RMSE rises to 0.08–0.10 mph. This confirms the theoretical expectation that CS performance hinges on signal sparsity.
• Communication Savings: Compression reduces transmitted data by 80 % for CR = 0.2. In LTE and early‑5G field tests, the average end‑to‑end latency for a compressed packet is under 30 ms, comfortably meeting the latency budgets of most V2X safety applications.

Computational Overhead
The on‑board compression requires only a few milliseconds per second of data, leaving ample CPU headroom for other vehicle functions. Reconstruction on a server equipped with a multi‑core Xeon processor completes a 10‑second window (≈100 samples) in under 0.2 seconds using OMP, demonstrating scalability for city‑wide deployments.

Implications and Future Work
The study validates that real‑time CS can be a practical solution for the “big data” challenge in connected‑vehicle ecosystems. By exploiting the natural smoothness of speed trajectories, CS achieves high compression without sacrificing the fidelity needed for downstream analytics such as traffic flow estimation, incident detection, and eco‑driving feedback. The authors suggest several avenues for further research:

1. Multi‑modal CS – extending the framework to jointly compress acceleration, yaw rate, LiDAR point clouds, and other sensor streams, potentially using joint‑sparsity models.
2. Adaptive Measurement Matrices – designing Φ that adapts to detected non‑sparse events (e.g., sudden braking) to maintain reconstruction quality.
3. Deep‑Learning‑Assisted Reconstruction – leveraging convolutional or recurrent neural networks to improve recovery in highly dynamic scenarios.
4. Security & Privacy Integration – embedding lightweight encryption within the CS measurement process to protect driver privacy while preserving compressibility.
5. Large‑Scale Field Trials – deploying the system across multiple CV fleets to quantify network‑wide bandwidth savings and to evaluate the impact on real‑time traffic‑management platforms.

Conclusion
The paper demonstrates that a carefully engineered compressive sensing pipeline can compress CV speed data by a factor of five (or more) in real time, transmit it over existing V2X links with negligible latency, and reconstruct the original signal with sub‑0.1 mph error. This approach offers a scalable, cost‑effective pathway to harness the full potential of connected‑vehicle data for safer, more efficient, and more sustainable transportation systems.