Abstract
This study develops a deep multimodal learning framework for lane-level traffic-speed estimation using five sensing sources: vehicle detection systems (VDS), closed-circuit television (CCTV), light detection and ranging (LiDAR), remote traffic microwave sensors (RTMS), and automatic vehicle classification (AVC). The framework integrates sensor-specific features, road attributes, modality availability, and spatial relationships. Using traffic data from Siheung, Republic of Korea, the proposed model was evaluated against five single-modal baselines. The multimodal model achieved a weighted mean absolute percentage error (WMAPE) below 10% across evaluated time steps, outperforming single-modal models in stability and cross-modal transferability. Improvements were more pronounced in selected multi-lane environments and areas with limited sensing coverage. In route guidance, lane-level multimodal information reduced mean minimum travel time by 7.31% compared with link-level information. These findings demonstrate the potential of multimodal integration for accurate lane-level estimation, recovery of missing sensor observations, and improved route guidance.