An improved binocular LSD_SLAM method for object localization

Abstract

Vision sensors can simulate human eyes to process visual scenes. Vision-based object localization technology has been widely studied and applied in some fields such as autonomous driving. SLAM technology can estimate the surrounding environment and locate itself in real time without prior knowledge. The existing binocular vision positioning SLAM has problems such as insufficient positioning accuracy, high data requirements and high computational cost. This paper proposes an improved binocular LSD_SLAM method with census transform, which purpose is to reduce the requirement for the initial value and optimize the localization accuracy. Experiments in several office scenarios show that the proposed method is improved on Average Precision and Average Recall, and its computing cost still needs to be improved.