Project Overview
This project was developed in collaboration with Tainan City Government and Far EasTone Telecom as part of a smart-city initiative to improve pedestrian flow management at tourist attractions and transportation hubs. The system combines GRU-based time-series prediction with geospatial analysis to forecast pedestrian density up to 2 hours ahead, visualized through an interactive web dashboard.
Telco-derived crowd density data (anonymized mobile signal counts) was fused with weather, calendar, and event data to train the forecasting model, achieving strong results even during irregular holiday and event spikes.
Key Features
- GRU (Gated Recurrent Unit) sequence model with multi-variate input: crowd density, weather, time features, event flags.
- Geospatial heatmap overlays on Leaflet.js web dashboard showing predicted density by zone.
- Real-time ingestion pipeline from Far EasTone's anonymized telecom crowd signals.
- Achieved MAE below 8% on held-out test data including holiday seasons.