Dynamic visual SLAM based on probability screening and weighting for deep features

Most Simultaneous Localization and Mapping (SLAM) systems highly rely on static environments assumption, leading to low pose estimation accuracy in dynamic environments. Dynamic Visual SLAM (VSLAM) methods have exhibited remarkable advantages in eliminating negative effects of dynamic elements. However, most current methods, only built on traditional indirect VSLAM using hand-crafted features, are still inadequate in utilizing and processing deep features. To this end, this paper proposes a dynamic VSLAM algorithm based on probability screening and weighting for deep features. Specifically, a deep feature extraction module is designed to generate deep features leveraged in the overall pipeline. Then, probability screening and weighting scheme is proposed for processing deep features, through which the dynamic deep feature points are eliminated in a coarse-to-fine manner and the various contributions of static ones is distinguished. Sufficient quantitative and qualitative experiments prove that our proposed method is superior to other counterparts in terms of localization accuracy.