Multi-view Visual-inertial Fusion for Precise AGV Navigation in Workshops

Aiming to achieve precise navigation of AGVs (automated guided vehicles) in workshops, a real-time multi-view localization method with tightly-coupled visual-inertial fusion is proposed. This method address the challenges faced by existing visual-inertial fusion based localization methods, such as lack of absolute pose, inaccurate estimation of absolute scale, and significant cumulative errors. Firstly, a visual-inertial fusion based AGV navigation framework with global consistency is designed to establish a global reference coordinate system and achieve long-term drift correction. Next, a multi-view camera and IMU (inertial measurement unit) joint initialization method is proposed to tackle the issue of inaccurate scale estimation in the initialization phase of visual-inertial fusion. This method utilizes a maximum posteriori probability model to obtain more accurate initialization parameters. Furthermore, a QR (quick response) code based pose correction model is proposed to compensate some keyframes periodically, and thus mitigating the effects of error accumulation and inertial deviation drift in the tracking estimation part. In addition, a pose constraint optimization model is proposed to address the issue of local extremum in the optimization mapping part and improve AGV localization accuracy. Finally, the proposed method is validated on the constructed AGV navigation platform in a workshop and compared against state-of-the-art visual-inertial navigation methods. The results demonstrate the superiority of the proposed method in terms of time efficiency and positioning accuracy. Specifically, translation RMSE (root mean square error) is less than 50 mm, and rotation RMSE is less than 2°.