GPU-accelerated numerical differentiation for loop closure in visual SLAM

This thesis introduces a technique that leverages a GPU to enhance the efficiency of loop closure in visual-inertial SLAM systems, particularly in the approximation of Jacobians using the Finite Difference Method (FDM). Traditional FDM techniques often encounter computational overhead due to repeated perturbations in pose graphs. This work addresses this challenge with a novel methodology that includes strategic graph partitioning and an optimized approach to Jacobian approximation. By integrating this technique into ORB-SLAM3's g2o framework, the linearization process is significantly enhanced. The evaluation of this approach, conducted on 12 sequences of varying lengths from the EuRoC and TUM-VI datasets, demonstrates a speedup of up to 4.23x in the linearization stage and an overall performance improvement of up to 2.08x in the optimization process.