This work addresses ego‑trajectory prediction for autonomous vehicles operating in urban road construction zones. Temporary changes in lane geometry and connectivity, together with interactions with workers, make future vehicle motions hard to anticipate. To tackle this, we built WZDrive-Data, a multimodal dataset comprising 1,183 driving segments collected from 33 participants in a driving simulator that integrates prerecorded worker animations and VR‑controlled workers. The dataset records driver control commands, gaze signals, vehicle trajectories, and scene metadata. Leveraging these signals, we propose a multimodal ego‑trajectory predictor named WZDrive, which fuses control and gaze information. On our benchmark, WZDrive achieves a mean minADE (K=6) of 1.09 m and minFDE (K=6) of 2.12 m, improving over the strongest baseline by 7.18% and 13.18% respectively. The study fills a gap in autonomous‑driving research on construction zones and demonstrates the value of human‑in‑the‑loop data for multimodal learning.
Review