Construction and Development of Visual Simultaneous Localization and Mapping System for Multi-person Dynamic Spaces

Abstract

In this study, we developed a visual simultaneous localization and mapping (SLAM) dynamic object filtering system applicable to multi-person dynamic spaces. When a visual mobile vehicle with an edge computing platform is used in the dynamic living space, the visual SLAM system will likely be disturbed by dynamic persons, leading to poor mapping and localization. We also implemented the dynamic object filtering function by preprocessing with a semantic segmentation neural network and designed a high-performance neural network architecture for the edge computing platform. The proposed design is capable of real-time operation to improve the effect of dynamic objects on the visual SLAM system.