Emergency Incident Detection from Crowdsourced Waze Data Using Bayesian Information Fusion

A principled Bayesian framework that fuses noisy, redundant crowdsourced Waze reports across space and time to detect and localize traffic accidents. On Nashville data it beats count/reliability baselines in F1 and AUC, and on average flags accidents 5.92 minutes before the official record.

Problem

Emergency services are reactive: they wait for 911 calls. Crowdsourced platforms like Waze offer a proactive signal, but reports are unreliable (variable user credibility), spatio-temporally uncertain (people report after driving past), and redundant (many reports per incident; Fig. 1). Prior work used simple aggregation (counts or averages) of surface features with no principled treatment of uncertainty, and did not analyze the effect of spatio-temporal resolution.

Framing: humans act as unreliable sensors and accidents act as faults, which makes this an analogue of fault diagnostics in cyber-physical systems.

Fig. 1: Waze reports (blue) vs. actual incidents (red) in the same region at two different times, which shows the fusion challenge.Figure from Senarath et al., IEEE/WIC/ACM WI-IAT 2020.

Approach

1. Spatio-temporal discretization: H3 hexagonal grid regions; fixed time steps (Fig. 3).
2. Grouping reports to incidents: segmentation (collect all reports in a region over the incident period) vs. DBSCAN clustering on location and time (best ε = 0.8, silhouette 0.8156).
3. Incident detection: Naive Bayes posterior P(incident | reports), treating each report’s reliability/10 as its true-positive probability; priors from historical E-TRIMS counts per region and hour (Eqs. 2–5).
4. Incident localization: each report covers a circle of radius δ = v·t_r (reporter speed × reaction time); the likelihood is proportional to its area overlap with each region (Eqs. 6–8).
5. Sequential updating: the posterior at step m becomes the prior at m+1; new reports join existing clusters or start new ones (Sec. IV-G).
6. Alert decision: logistic regression / random forest learns the probability threshold against E-TRIMS ground truth (Sec. IV-H).

Fig. 3: Pipeline across time steps T_m → T_m+1. Red boxes are data and clustering, green are detection and localization, purple is the alert decision.Figure from Senarath et al., IEEE/WIC/ACM WI-IAT 2020.

Key results

• Evaluation: 5-fold CV; Oct–Dec 2019; 33,218 Waze reports, 2,878 E-TRIMS incidents.
• Best hyperparameters: incident period T′ = 25 min, time step = 1 min, δ = 100 m, H3 resolution 6.
• Best F1: 48% (M6, segmentation-based plausibility + logistic regression) vs. 45% for the best baseline (M2). Best AUC: 71% (M7/M8, all features, clustering) vs. 65% baseline (Table II).
• The conclusion reports a relative F1 gain of over 5% against baselines.
• Clustering gives higher precision but lower recall; the random-forest baseline has the highest recall (78%) but low precision.
• Early detection: the best model (M6) flags accidents on average 5.92 min before the official E-TRIMS record. In the case study (Fig. 8), Waze reports arrive 12 min before the official record.

Contributions

1. A novel Bayesian-theoretic method for spatio-temporal fusion of noisy crowdsourced reports, with uncertainty quantified as calibrated probabilities (confidence per prediction).
2. Sequential updating of incident probabilities as new reports arrive (real-time).
3. Analysis of how spatio-temporal resolution affects detection.
4. An extensible framework for other noisy crowdsourced sources.

Limitations & future work

• Matching crowdsourced reports to formal records is hard because of human error and reporting delays.
• Assumes constant reporter speed and independent reports.
• Future work: automated feature extraction and fusion; road-segment fidelity instead of regional grids.

@inproceedings{Senarath2020b,
  author       = {Senarath, Yasas and Nannapaneni, Saideep and Purohit, Hemant and Dubey, Abhishek},
  booktitle    = {2020 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT)},
  title        = {Emergency Incident Detection from Crowdsourced Waze Data Using Bayesian Information Fusion},
  year         = {2020},
  organization = {IEEE},
  pages        = {187--194},
  keywords     = {Conference,first,Crowdsourced},
  url          = {https://ieeexplore.ieee.org/document/9457694/}
}