Enhancement of Visual SLAM Precision Using Semantic Image Segmentation for Automotive Systems

Abstract

In the field of advanced driver assistance systems (ADAS) testing and autonomous driving (AD) feature evaluation, novel approaches relying on augmented reality (AR) promise to deliver cost-saving benefits. These forward-looking approaches leverage vSLAM techniques to create mapping solutions that are essential for augmentation. A critical challenge, however, is maintaining the high precision required for these maps and, by extension, the SLAM algorithm itself. This precision is often compromised by the presence of false-positive detections of feature points. In response to this challenge, this paper presents an improvement to the ORB-SLAM3 algorithm. The proposed approach incorporates semantic segmentation without compromising processing speed to increase the precision and reliability of the SLAM system. This is to ensure that the integration of AR-based solutions in the automotive sector is both effective and sustainable, providing tangible benefits in the testing and development of ADAS and autonomous driving technologies.