In context
By 2024, visual SLAM had become a core enabler for indoor mobile robots, yet its reliability still degraded sharply in feature-poor environments such as white walls and plain floors. This limitation directly affected automation tasks like AGV navigation and service robot localization, prompting research into fusing learning-based feature matching with classical geometric estimation.
What was reported
Researchers at China Jiliang University proposed Lo-RINS, a visual-inertial SLAM system built on VINS-RGBD, to improve robustness in indoor weak texture scenes. The front end integrates a deep learning module based on attention mechanisms (LoFTR) to directly match features between adjacent frames, supplementing traditional Harris corner detection when insufficient features are extracted. A downsampling step using Gaussian kernels reduces computational load before matching.
A camera depth confidence model was introduced to weight feature points based on their depth and pixel position, mitigating drift from distant features. The confidence values also serve as piecewise thresholds in a dynamic robust kernel function that optimizes bundle adjustment in the back end.
In real-scene experiments, the algorithm demonstrated significant robustness in weak texture environments. Compared with the VINS-RGBD baseline, the absolute trajectory error was reduced by 50.38%, and the relative trajectory error by 85.75%.
Why it mattered
This work showed that combining learned feature matching with explicit depth uncertainty modeling can substantially improve visual SLAM accuracy in challenging indoor settings, offering a practical path toward more dependable robot localization in real-world manufacturing and service environments where textureless surfaces are common.
Compared with VINS-RGBD algorithm, the absolute trajectory error of the proposed algorithm is reduced by 50.38% and the relative trajectory error is reduced by 85.75%.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-05-08 · “一种室内弱纹理环境下的视觉SLAM算法”