Weak-Supervised Simultaneous Panoramic View Generation and Articulation Angle Estimation for Long Combination Vehicles

Panoramic view generation and articulation angle estimation are essential for enabling active safety and control in complex dynamic systems such as long combination vehicles (LCVs). Traditional approaches rely on mechanical sensors for angle measurement, which are susceptible to noise and require frequent maintenance, potentially degrading lateral control performance and causing misalignment in panoramic views. Moreover, panoramic view generation for LCVs remains underexplored due to their extended length and complex articulated structures. To address these challenges, this study proposes a real-time, weak-supervised framework with offline training and online inference modules to achieve simultaneous panoramic view generation and articulation angle estimation. In the offline stage, the angle estimation network is trained using noisy angle data as weak supervision, guided by image-level self-supervision signals like photometric consistency. In the online stage, the network directly infers articulation angles from raw fisheye image pairs, eliminating the need for preprocessing or computationally expensive online optimization. Field experiments on a three-carbody-articulated LCV validate the effectiveness of the framework in accurately estimating articulation angles and generating holistic, color-balanced panoramic views under varying conditions. The system addresses information loss and photometric inconsistency, especially during large-angle turns, offering a practical, sensor-independent vision solution to enhance safety and deployment flexibility in complex vehicle systems.