Enhanced Neural SLAM with Semantic Segmentation in Dynamic Environments

Abstract

Mobile robots face a significant challenge when navigating in dynamic environments such as human crowds. Existing research works typically employ separate data or simulators for dynamic object detection and navigation tasks. This paper combines the photo-realistic simulator Habitat with embedded dynamic objects to create a playground for optimizing vision-based navigation algorithms. To validate our system, we implement and train three approaches - zero masks, image memory inpainting, and semantic map filter - using semantically segmented information. We then implement and evaluate an adapted reinforcement-learning-based SLAM algorithm using the validation Gibson dataset. The results indicate that moderate localization bias occurs when the environment is small and the navigation process is short. However, the error will accumulate over time. Additionally, a dynamic environment leads to less accurate localization. All three approaches can reduce the localization error, while the semantic map filter approach shows the best overall performance.