Adaptive calibration and spatio-temporal filtering for robust highway wrong-way driving detection

Abstract

Wrong‐way driving on highways is a low‐frequency but high‐risk event that requires reliable detection by future transportation systems and roadside intelligent infrastructure. Existing camera‐based solutions often rely on fixed lane‐direction settings or short‐term trajectory displacement, making them vulnerable to camera viewpoint changes, vehicle occlusion, bounding‐box drift, and stationary vehicles with flashing lights. This study proposes an adaptive calibration and spatio‐temporal filtering framework for robust detection of highway wrong‐way driving under operational surveillance disturbances. The framework first derives lane-level reference directions from road semantic segmentation and historical vehicle trajectories, and then uses lightweight difference hashing to perceive large viewpoint rotations and update direction references online. Candidate wrong‐way trajectories are subsequently verified using a cascaded strategy that integrates temporal continuity filtering, perceptual difference map analysis, and max‐margin adaptive boundary matching. A dedicated highway surveillance dataset containing 99 video sequences was constructed from real-world open‐highway, tunnel, and camera-rotation scenarios, together with a feature-difference dataset for training the secondary verification module. The proposed framework achieved 97.62% precision, 95.35% recall, and a 96.47% F1‐score, while maintaining a low operational false‐alarm rate against representative trajectory‐ and tracking‐based baselines. These results indicate that adaptive infrastructure-side perception can reduce false alarms and support proactive highway traffic safety management in future transportation systems.