In high-end equipment manufacturing, aerospace, shipbuilding, and other industrial fields, tasks such as precise localization of industrial robots, assembly of large components, and target docking rely heavily on the ability to obtain real-time six-degree-of-freedom (6DoF) pose information. Visual measurement methods have been widely used in simultaneous localization and mapping (SLAM) due to their non-contact, low power consumption, and rich information acquisition characteristics. However, existing visual SLAM algorithms based on natural features can easily suffer from tracking interruption and accumulated errors when facing texture feature loss. Although some researchers have improved the robustness of the system by introducing artificial planar markers, it is still difficult to meet the high-precision measurement requirements in industrial environments. To address these issues, we introduce industrial high-reflective markers to replace natural features, providing visual observation information and improving resistance to environmental interference, dynamic stability, and measurement accuracy. Based on the introduction of industrial reflective features, we focus on high-precision global map construction and high-precision real-time localization to achieve more accurate and stable pose estimation.

To achieve high-precision real-time localization in unstructured industrial environments, we introduced industrial directional high-reflective markers. By recognizing and extracting the centers of encoded features, we obtained high-precision visual feature data and decoded the feature IDs to enable feature matching between frames. Recognizing that the localization of reflective features can be easily disturbed by manual factors, resulting in uneven distribution, we improved the accuracy and stability of global localization. With directional reflective markers as observation features, we divided the entire measurement process into two stages: map construction for reconstructing a high-precision map and 6DoF real-time pose measurement for recovering high-precision poses. The global a priori information from the former stage provided global auxiliary constraints for the latter, enabling more accurate and stable pose estimation. During the rapid construction of the global map, we relied on the visual sensors to fully observe the reflective features in the environment, performing pose estimation and initial map construction of the reflective features simultaneously. To improve the real-time efficiency of the system while maintaining high accuracy, we optimized the distribution of the key frame network structure to select the best key frames. External constraint information was utilized to introduce global scale information, and global optimization was performed based on bundle adjustment (BA). In the visual-inertial real-time localization section, we integrated the visual sensor with an inertial measurement unit (IMU). The IMU provided an initial pose estimate, ensuring continuity in areas where reflective encoded features were absent. We utilized the pose of key frames and map point information from the global information as global a priori constraints. These constraints were combined with the current image frame containing common observation points for tightly coupled visual-inertial joint optimization. Throughout this process, the map was updated with the latest observations.

To verify the constraint effect of the improved key frame selection strategy on the map in this paper, we use the map points obtained after BA optimization with all images as the measurement benchmark to analyze the three-dimensional coordinate accuracy of the generated global map points. At the same time, we compare the method in this paper with the image network design (IND) method in Ref. [19] to verify the impact of the improved method (Fig. 10, Table 2, and Table 3). The results show that the proposed method improves the translation accuracy by 25.29% compared to the method in Ref. [19], reduces the maximum outliers by 64.72%, and decreases the proportion of bad points with an error greater than 1 mm by 4.74%. To validate the localization accuracy of the designed system in this article, we use the T-Mac 6DoF measurement device of the laser tracker as the comparison benchmark. We also verify that after adopting reflective features, ORB-SLAM3 improves its accuracy by 74.5% compared to natural features (Fig. 11 and Table 4). Subsequently, we compare the proposed method with ORB-SLAM3 using reflective features and the PnP based on the global map in terms of accuracy through four sets of experimental data. The results indicate that the proposed method outperforms ORB-SLAM3 using reflective features and the PnP algorithm by an average of 66.72% and 12.93% (Table 5) in localization accuracy, respectively. The absolute trajectory errors of the experimental results are all less than 2 mm, and the relative attitude errors are less than 0.03° (Table 6), achieving high-precision real-time localization in unstructured industrial environments.

Against the backdrop of high-precision real-time localization in unstructured industrial environments, we propose a visual SLAM method based on industrial high-reflectance features. This method employs optimal network optimization to select a certain number of best key frames and performs global optimization on the selected key frames and encoded map points to obtain a global a priori map. During subsequent real-time localization, real-time global pose estimation is carried out based on global a priori information and inertial odometry information, and the confidence of each map point is assigned through an information matrix. More accurate map point information is obtained through continuous updating during subsequent localization and fed back to the 6DoF pose. Finally, experimental results are analyzed based on the T-Mac benchmark. Under the assistance of global information, the estimated pose of the proposed method exhibits better localization accuracy and robustness compared to the ORB-SLAM3 algorithm and PnP algorithm using reflective features.