Detecting and Classifying Traffic Rule Changes from Induction Loop Data

Deep learning for smart city map maintenance: detecting road blockages and traffic sign changes from urban sensor data - without cameras or GPS traces.

📄 Published paper: IEEE Xplore · Proceedings of the 2023 IEEE International Conference on Big Data (BigData 2023)

Overview

Keeping digital road maps accurate for navigation apps, autonomous vehicles, and smart city systems usually depends on expensive methods like satellite imagery or street-view cameras. This project explores a cheaper, privacy-preserving alternative: using existing induction loop sensors (already embedded in roads for traffic management) to automatically detect and classify changes in traffic rules.

We designed a deep learning pipeline (based on a modified U-Net with spatio-temporal components) that takes in city-wide traffic flow data (vehicle speed and road occupancy) and predicts:

• Road or lane blockages
• Speed limit changes
• New turn restrictions (left, right, U-turn, straight-only)

The model was evaluated on both simulated (SUMO/LuST, Luxembourg) and real-world (HERE Traffic API) data, achieving F1-scores above 80% in detection and over 75% in multi-class classification, despite challenges like sparse sensor coverage, high-dimensional input, and severe class imbalance.

Why this matters

• Cost-effective: no need for camera fleets or manual surveying, induction loops are already deployed in most cities
• Privacy-preserving: unlike GPS traces from taxis/buses, this approach doesn't track individual vehicles
• Applicable to: autonomous driving map maintenance, smart city infrastructure monitoring, urban traffic anomaly detection, digital twin systems

Method overview

Illustration of how induction loops detect passing or stationary vehicles.

Sliding‑window data processing and prediction workflow.

Pseudocode

For a structural walkthrough of the pipeline from tensor construction through the loss function, see code/pseudocode.md.

Quick preview

1. Input tensor construction
2. Multi-pooling down-sampling
3. 3D overlapping sliding windows + augmentation
4. Model: 3D U-Net + stateful ConvLSTM + residual connections
5. Class-imbalance-aware weighted loss
6. Confidence-weighted prediction aggregation

Minimal Demo

A small, from-scratch PyTorch demo code/demo_convlstm_synthetic.py illustrating the ConvLSTM-based spatio-temporal detection idea from the paper, trained on synthetic data. Not the original implementation or dataset - see the file's docstring for details.

Key results

Experiments on simulated data

Experiments on real-world data

Achieved F1‑scores above 80% in detection and over 75% in multi‑class classification, validated on both simulated (SUMO/LuST, Luxembourg) and real‑world (HERE Traffic API) datasets.

Research Talk Slides

To make the paper more accessible, here are slides from an internal lab presentation summarizing the work: Download Full Slides (PDF)

Citation

If you find this work useful, please cite:

BIBTEX

@inproceedings{zhanabatyrova2023detecting,
  title={Detecting and Classifying Changes in Traffic Rules using Induction Loop Data},
  author={Zhanabatyrova, Aziza and Leite, Clayton and Xiao, Yu},
  booktitle={2023 IEEE International Conference on Big Data (BigData)},
  pages={1248--1255},
  year={2023},
  organization={IEEE}
}

APA

Zhanabatyrova, A., Leite, C., & Xiao, Y. (2023, December). Detecting and Classifying Changes in Traffic Rules using Induction Loop Data. In 2023 IEEE International Conference on Big Data (BigData) (pp. 1248-1255). IEEE.

Keywords

traffic anomaly detection change detection IoT deep learning U-Net ConvLSTM spatio-temporal data smart city urban sensing induction loop sensors traffic sign detection digital map maintenance map updating digital twin autonomous driving HD maps intelligent transportation systems traffic flow prediction class imbalance tensor-based learning multi-label classification SUMO traffic simulation sim-to-real transfer domain adaptation privacy-preserving sensing HERE Traffic API

❓ FAQ

1. Why induction loops? They’re already deployed in most cities and don’t track individuals.
   

2. Can this run in real time? The current pipeline processes data in offline batches; the paper's Section V-D discusses potential directions like handling transient change dynamics and closing the simulated-to-real domain gap.

3. Could this run in a distributed/edge setup instead of centrally? The model's spatial decomposition into overlapping sliding windows (Section III-B) maps naturally onto distributed processing. Each city region could be handled by a local edge node, exchanging only the boundary/overlap information needed to preserve cross-region context, rather than streaming all induction loop data to one central server. The current implementation runs centrally on a single GPU; distributed deployment is a promising extension, not something evaluated in the paper.

📬 Contact

Google Scholar https://scholar.google.com/

Email: zhanabatyrova@gmail.com

LinkedIn https://www.linkedin.com/

---

This repository accompanies the published paper above. For questions about the method or data, feel free to open an issue.