GSL-VO: A Geometric-Semantic Information Enhanced Lightweight Visual Odometry in Dynamic Environments

Abstract

Recently, learning-based Visual Odometry (VO) has attained remarkable success in vision-based measurement, especially in indoor robotics. Unfortunately, existing methods usually under-explore geometric-semantic information, thus resulting in inefficient perception in unseen dynamic environments. Meanwhile, they are usually time-consuming, since they typically rely on high-complexity semantic segmentation models, resulting in concurrency reduction. In this paper, we develop a geometric-semantic information enhanced lightweight visual odometry (GSL-VO) that can work particularly well in dynamic environments. Specifically, on the one hand, to improve the robustness of VO through geometric-semantic information, we first come up with a novel image enhancement module to tackle motion blur, thus enabling accurate geometric and semantic information extraction. Second, we design an adaptive geometric-semantic information processing module that combines geometric and semantic information to retain reliable features for pose measurement. Moreover, semantic information is expressed via a probability framework for accurate and robust movable object extraction. On the other hand, we further propose a lightweight semantic segmentation model that enjoys an efficient multi-level feature aggregation capability to address the speed bottleneck of VO. A series of experiments on two well-known RGB-D dynamic datasets indicate that our proposed method is both accurate and fast: while achieving a significant average improvement of 70.5% in ATE over state-of-the-art learning-based VO on Bonn RGB-D Dynamic dataset, GSL-VO leads to high 22.3 FPS on a low-cost platform, which makes it well-suited for practical scenarios. Remarkably, on a challenging dynamic sequence of TUM RGB-D dataset, GSL-VO improves the baseline VO by 88.9% in ATE.