A Practical Roadside Multi-View Multi-Sensor Calibration Framework Based on Monocular Vehicle Localization

Multi-view multi-sensor calibration (MVMSC) in roadside scenarios is a prerequisite for the fusion of multi-source information obtained by various sensors at different locations.However, the current roadside MVMSC methods based on the cascading spatial transformation (CST) can lead to considerable cumulative errors when deployed on a large scale.And traditional camera calibration in roadside MVMSC is insufficient to handle scenarios with changing camera parameters and requires significant manual intervention when deployed extensively.Additionally, existing methods are often limited to controlled or single-view scenarios, lacking consideration for various practical problems that arise in large-scale deployments.To address these challenges, this research proposes a customized roadside MVMSC framework based on monocular localization in a non-CST manner.This framework directly transforms all sensor coordinate systems into the global coordinate system, thus mitigating the accumulation of errors and minimizing the need for extensive manual intervention associated with CST.Furthermore, For achieving precise roadside monocular global localization while considering scenarios with varying camera parameters and unknown camera heading angles, this research incorporates deep learning into camera calibration, distance and angle estimation, and automatic heading angle calculation, reducing manual intervention.Finally, in mitigating the potential reduction in calibration accuracy in real-world settings, this paper utilizes geolocation cues and an optimization algorithm based on stochastic gradient descent to improve MVMSC precision.The proposed method has been systematically tested in real-world scenarios under various parameters and data conditions.Experimental results demonstrate its efficiency and accuracy advantages over previous CST-based modes.Compared to previous manual methods based on CST, our approach reduces operation time by approximately 89% and improves accuracy by about 91%.