An Enhanced Trajectory Optimization Approach for Equirectangular Camera-based Visual SLAM

Compared to other cameras, panoramic cameras offer the advantages of rich feature points and enhanced robustness in tracking texture-deficient regions within visual SLAM solutions. However, the accuracy of their estimated trajectories can be significantly affected by invalid feature points. To mitigate this influence, this paper proposes a trajectory optimization method based on OpenVSLAM incorporating object detection and mask filtering techniques. Specifically, we introduce a mask-generated network to effectively filter out invalid feature points from the captured images. Additionally, an object detection mechanism is implemented to enable real-time rejection of dynamic character feature points through socket communication. Experimental results obtained in various site scenes demonstrate that the proposed algorithm improves trajectory accuracy and validates the effectiveness of our approach.