DVDS: A deep visual dynamic slam system

Abstract

Simultaneous localization and mapping (SLAM) utilizing visual sensors represent an extensively investigated research area, holding significant potential for advancements in robotics and autonomous vehicular systems. Recently, dense SLAM systems underpinned by learning-based methodologies have showcased superior accuracy and robustness compared to conventional techniques. Nevertheless, contemporary learning-based SLAM systems exhibit notable discrepancies in pose estimation, particularly within dynamic environments. In addition, the constrained receptive field of convolutional features in these methods impedes their efficacy when confronted with homogeneous, texture-less images, rendering them vulnerable to noise perturbations. We develop a novel deep visual dynamic slam (DVDS) system that exploits solely static pixels within images to retrieve the camera poses. Specifically, we formulate a dynamic object exclusion mechanism that excises dynamic constituents within the scene before the optical flow computation, thus optimizing the precision of the estimation. In addition, we unveil an efficient dispersive transformer (DisFormer) that facilitates per-pixel features in assimilating long-range information from surrounding features, culminating in constructing more precise 4D correlation volumes. Building on the DisFormer, we suggest a Disformer-based gated recurrent unit (GRU) to generate a refined flow field coupled with a confidence map, which is subsequently employed by the dense bundle adjustment layer to iteratively rectify the residuals of inverse depths and associated camera poses. The global receptive field provided by the DisFormer promotes information integration from a wider contextual window, thus improving the robustness of our SLAM system. Comprehensive experiments underscore that our proposed DVDS system manifests superior efficacy compared with state-of-the-art works across both static and dynamic scenes.