Background Mobile robot applications rely heavily on simultaneous localization and mapping (SLAM) ; however, visual SLAM systems often struggle to maintain accuracy and resilience in dynamic environments. A key challenge is the ineffective filtering of dynamic feature points, which leads to localization errors. Recent advancements have introduced optical-flow techniques to remove moving objects, but further improvements are needed to enhance both accuracy and efficiency. Methods This research improves dynamic object filtering by introducing a region-based convolutional neural network (R-CNN) to eliminate highly dynamic objects. A cross-layer feature pyramid network (CFPN) is integrated to enhance feature extraction in dynamic scenarios. Additionally, the elite moth–flame optimization algorithm (EMFOA) is employed alongside optical-flow tracking to refine feature-point matching. The approach optimizes a classical optical flow objective across discrete grids, explicitly targeting error criteria to improve flow field quality. By leveraging the structured mapping space, computational complexity is reduced from quadratic to linear. Results Compared to existing methods, the proposed technique demonstrated superior precision. On the SittingXYZ dataset, it achieved the lowest absolute pose error (APE): mean = 0. 2211, median = 0. 2359, RMSE = 0. 2159, and standard deviation = 0. 0716. On the Sitting dataset, the classifier reduced APE mean values by 39. 90%, 33. 96%, 29. 92%, 24. 44%, 16. 60%, and 8. 56% compared to DGS-SLAM, YOLO-SLAM, ORB-SLAM2, OPF-SLAM, DynaTM-SLAM, and DI-SLAM, respectively. These results underscore the enhanced localization precision and robustness of the proposed system in continuously evolving environments.
R et al. (Fri,) studied this question.