Presentation Information
[3K2-OS-27b-04]Traffic Estimation with Task-Aware Clustering of Multimodal Geo-Semantic Features
〇Shuntaro Masuda1, Itsuki Hirai2, Toshihiko Yamasaki1 (1. The University of Tokyo, 2. MD Inc.)
Keywords:
Multimodal Unstructured Data,Traffic Volume Estimation,Geo-Semantic Representations,Task-Aware Clustering,Real Estate Analysis
Estimating automobile traffic volume at the road-segment scale is important for real estate location and trade area analysis. However, publicly available traffic data are limited to sparse observation points, and surrounding data usable as explanatory variables are also insufficient, while acquiring such data often involves high costs.
This study proposes a traffic volume estimation method using three types of publicly available multimodal unstructured data: satellite imagery, street view imagery, and road names. Large multimodal models are employed to extract and integrate semantic features related to road information and roadside land use, enabling the use of rich geo-semantic information beyond conventional approaches. Furthermore, task-aware clustering is introduced to aggregate integrated road features into cluster features suitable for regression. Experiments with traffic census data show that both multimodal feature integration and the addition of cluster features contribute to improved predictive performance. In the best XGBoost setting, MAPE improved by 2.83%. Fine-grained geo-semantic features also expand variables available for interpretability analysis and support qualitative evaluation of roads and locations.
This study proposes a traffic volume estimation method using three types of publicly available multimodal unstructured data: satellite imagery, street view imagery, and road names. Large multimodal models are employed to extract and integrate semantic features related to road information and roadside land use, enabling the use of rich geo-semantic information beyond conventional approaches. Furthermore, task-aware clustering is introduced to aggregate integrated road features into cluster features suitable for regression. Experiments with traffic census data show that both multimodal feature integration and the addition of cluster features contribute to improved predictive performance. In the best XGBoost setting, MAPE improved by 2.83%. Fine-grained geo-semantic features also expand variables available for interpretability analysis and support qualitative evaluation of roads and locations.