From augmentation to inpainting: improving Visual SLAM with object detection and removal, signal enhancement techniques and GAN-based image inpainting

Abstract

In dynamic indoor environments, the operation of Visual Simultaneous Localization and Mapping (vSLAM) systems requires careful consideration of moving objects, as they can significantly affect the stability of visual odometry and the accuracy of position estimation. This challenge requires the exploration of signal enhancement techniques to improve vSLAM performance. This thesis proposes a vSLAM system based on ORB-SLAM3 and YOLOR, augmented with the YOLOX object detection model, which achieves an improvement in accuracy of 2—- 4% compared to previous systems. By utilizing static feature points for camera position calculation and dynamic object tracking, the system effectively mitigates environmental disturbances. Furthermore, a novel approach is introduced that leverages Generative Adversarial Networks (GANs) for image inpainting following object removal, thereby enhancing both the accuracy and execution speed of the system. Through a comprehensive investigation, this study not only evaluates existing methods, but also proposes innovative denoising techniques. The integration of signal enhancement and advanced denoising contributes to improved accuracy, robustness, and computational efficiency in real-world vSLAM scenarios, thereby advancing the field’s capabilities.