Early Warning of Human Crowds Based on Query Data from Baidu Maps: Analysis Based on Shanghai Stampede

Jingbo Zhou, Hongbin Pei, Haishan WuAdvances 2017 · DOI · OpenAlex

Introduction

Predicting massive crowd gatherings is a critical component of public safety management, yet it remains notoriously difficult using traditional methods. The 2014 Shanghai Stampede, which resulted in 36 fatalities, highlighted the limitations of conventional crowd anomaly detection. Traditional systems rely heavily on video sensors and computer vision, which are often hindered by environmental noise, limited geographic coverage, and a lack of sufficient lead time—often providing warnings only minutes before an incident occurs.

This research introduces a novel early warning system that leverages digital intent to predict physical reality. By utilizing aggregated data from Baidu Map, the study demonstrates that spikes in digital routing queries can reliably predict physical crowd surges 1 to 3 hours in advance, providing authorities with a vital window for emergency intervention.

Methodology

The proposed system operates on the behavioral insight that individuals typically search for a destination on a map application before traveling. The authors developed a two-part framework to operationalize this data:

  1. Statistical Decision Method for Early Warning: Daily peak values of map queries are modeled using a log-normal distribution. An anomaly warning is triggered if current queries exceed a "warning line" defined by the historical mean plus a multiple of the standard deviation.
  2. Machine Learning for Density Prediction: To quantify risk, the researchers utilized a Gradient Boosting Decision Tree (GBDT) model to forecast the "positioning number" (physical crowd size) one hour into the future. The model incorporates 47 features, including historical positioning, recent map queries, and temporal context such as holidays and time of day.
System architecture for crowd density prediction and early warning using map query and positioning data.

Data and Empirical Analysis

The study utilizes proprietary, anonymized data from Baidu Map, focusing on two primary metrics:

  • Map Query Data: The volume of routing queries for a specific destination.
  • Positioning Data: The number of active users physically present in an area, serving as a proxy for population density.

The models were validated across six major Points of Interest (POIs) in China, including the Shanghai Bund and the Beijing National Stadium. A key finding is that digital intent consistently precedes physical presence.

Mutual information analysis showing the predictive power of map queries peaking at a 1-hour time lag.

As shown in the mutual information analysis, the peak correlation between query data and physical density occurs at a 1-hour time lag. This mathematically validates that knowing map query volume significantly reduces the uncertainty of predicting actual crowd size an hour later.

Key Results

  • Actionable Lead Time: The statistical method successfully triggered warnings 1 to 3 hours prior to actual crowd anomalies during high-profile events like New Year's Eve at the Shanghai Bund.
  • Predictive Accuracy: The GBDT model demonstrated high accuracy in tracking ground-truth spikes. Feature importance analysis revealed that the map query number from one hour prior is one of the most critical predictors of current density.
  • Ablation Insights: Removing map query features from the model significantly increased the Mean Absolute Error (MAE), proving that historical physical presence alone is insufficient for accurate forecasting without capturing user intent.

Conclusion and Future Work

This research demonstrates that digital footprints from map applications provide a powerful, low-cost, and high-lead-time alternative to traditional surveillance. While the model currently relies on a single proprietary data source, its success in predicting the dynamics of the Shanghai Bund suggests broad applicability for urban planning. Future work aims to expand the model to a wider variety of POIs and refine the threshold-based warning system to account for location-specific variances. The authors have already transitioned this research into a real-time early warning software system for large-scale deployment by city administrators.

Summary updated

BibTeX
@inproceedings{zhou2017early,
  author = {Jingbo Zhou and Hongbin Pei and Haishan Wu},
  title = {{Early Warning of Human Crowds Based on Query Data from Baidu Maps: Analysis Based on Shanghai Stampede}},
  booktitle = {Advances},
  year = {2017},
  doi = {10.1007/978-3-319-51929-6\_2},
}

Loading related content...

Figure