SCP-SLAM: Accelerating DynaSLAM With Static Confidence Propagation

Abstract

DynaSLAM is the state-of-the-art visual simultaneous localization and mapping (SLAM) in dynamic environments. It adopts a convolutional neural network (CNN) for moving object detection, but usually incurs a very high computational cost because it performs semantic segmentation using the CNN model on every frame. This paper proposes SCP-SLAM, which accelerates DynaSLAM by running the CNN only on keyframes and propagating static confidence through other frames in parallel. The proposed static confidence characterizes the moving object features by the residual defined by inter-frame geometry transformation, which can be computed quickly. Our method combines the effectiveness of a CNN with the efficiency of static confidence in a tightly coupled manner. Extensive experiments on the publicly available TUM and Bonn RGB-D dynamic benchmark datasets demonstrate the efficacy of the method. Compared with DynaSLAM, it enables acceleration by a factor of ten on average, but retains comparable localization accuracy.