An Integrated Vision-Sensor Fusion Framework for Wheel-Level Vehicle Trajectory Reconstruction and Axle-Load Estimation

Abstract

Accurate axle-load monitoring is essential for pavement design, evaluation and maintenance, yet conventional point-based pavement weigh-in-motion systems remain sensitive to uncertainties in actual wheel-load positions. To address this limitation, this study develops a physically interpretable vision–sensor fusion framework that reconstructs wheel-level loading geometry from monocular roadside video and integrates it with embedded pavement responses for axle-load estimation under in-service traffic conditions. Roadside video was processed through radial distortion correction, homography projection, YOLOv8-based truck detection, and HRNet-based keypoint identification. To infer tire-pavement contact points that are difficult to observe directly from an oblique monocular view, a surrogate keypoint strategy was introduced to reconstruct wheel-path trajectories on the road plane. Raw visual results were refined using rigid-body geometry, longitudinal kinematic constraints, non-holonomic motion priors, and Kalman filtering, converting image detections into temporally stable and physically consistent wheel-level trajectories. These trajectories were coupled with a simplified tire-pavement contact model and a multilayer-elastic-theory-based inversion model to estimate tandem rear-axle loads from embedded pavement responses. Field validation showed close agreement between vision-derived axle-passing times and sensor peaks, with an average timing discrepancy of approximately 8 ms. Reconstructed vehicle dimensions showed controlled errors mostly within ±0.15 m and identified rear-axle loads ranged from approximately 20 to 100 kN. Independent response validation reproduced measured stress at an unused sensor location, with a mean peak discrepancy of approximately 0.008 MPa, whereas estimation that disregarded the reconstructed wheel paths and assumed the load to act directly above the sensor produced markedly larger deviations. Results show that monocular traffic vision can provide mechanics-compatible wheel-level loading information for wheel-path-aware axle-load monitoring and reliable interpretation of in-service pavement sensor data.