MAS-DSO: Advancing Direct Sparse Odometry With Multi-Attention Saliency

Visual odometry (VO) is a critical component of simultaneous localization and mapping (SLAM) with extensive applications in robot navigation and beyond. However, prevalent VO methods often underperform in intricate environments with dynamic textures, insufficient lighting, and rapid rotational movements, primarily due to constrained feature selection and inadequate image structure comprehension. To address these challenges, this paper proposes a novel VO framework, termed Multi-Attention Saliency Direct Sparse Odometry (MAS-DSO) Specifically, MAS-DSO significantly bolsters performance and robustness through accurate recognition of visually salient regions and deep understanding of image structures. With regard to the problem of limited feature selection, we propose a Saliency Transformer Generative Adversarial Network (STRGAN) based on a multi-attention mechanism, narrowing the feature selection scope and enhancing its accuracy. Addressing the issue of limited understanding of image structure, we introduce a robust method for gradient computation to accurately determine the gradient values of features. Building on this, we have designed a dynamic gradient weight adjustment strategy that takes into account both the gradient magnitude and local image structure, thereby achieving precise gradient weight distribution. Comprehensive quantitative evaluations on the ICL-NUIM and TUM monoVO datasets reveal that MAS-DSO not only outperforms SalientDSO, DSO, and ORB-SLAM in performance metrics but also significantly surpasses other methods in saliency prediction performance and mapping quality. In conclusion, MAS-DSO not only augments feature selection efficiency but also enhances the processing prowess for diverse images in complex settings.