Guangfeng Liu, Zhuhua Hu, Yaochi Zhao, Ruoqing Li, Kunkun Ding, and Wenlu Qi
VSLAM, keyframe selection, motion model, bundle adjustment, DogLeg algorithm
Visual simultaneous localisation and mapping (VSLAM) is a technology that utilises visual sensors to achieve real-time localisation and map construction for robots in unknown environments. However, traditional VSLAM systems are prone to redundancy in keyframe selection, leading to excessive memory and computational burdens. Additionally, although bundle adjustment (BA) optimisation can enhance accuracy, it faces the challenge of high computational complexity. To address these issues, we propose a motion model-based keyframe selection method that effectively filters keyframes by analysing the relative motion between adjacent frames, reducing redundancy and enhancing the representativeness of keyframes. Simultaneously, we employ the DogLeg algorithm to replace the traditional Levenberg-Marquardt algorithm for local BA optimisation, achieving faster convergence and more efficient step length control. Experiments on the EuRoC and TUM VI datasets demonstrate that, compared to the traditional ORB-SLAM3, our method exhibits higher localisation accuracy and better real-time performance, especially showing superior performance in complex scenarios.
Important Links:
Go Back
