A Dynamic Visual SLAM System based on YOLOv5s and Scene Flow

The simultaneous localization and mapping (SLAM) system plays a vital role in autonomous driving and robotics. At present, numerous SLAM systems operate under the premise of a static environment. However, the presence of dynamic interferences frequently poses considerable difficulties for these systems, leading to issues such as data association inaccuracies and failures in tracking. This paper presents an innovative SLAM system that integrates YOLOv5s with scene flow techniques to effectively address challenges posed by dynamic interference. The system employs the YOLOv5s neural network for the precise detection of dynamic objects, subsequently utilizing geometric constraints of scene flow to remove dynamic feature points. Finally, the residual static feature points are employed for tracking purposes. We perform comprehensive experiments utilizing the TUM and Bonn datasets, and the results indicate that our approach markedly enhances both localization accuracy and robustness in comparison to ORBSLAM3. This finding substantiates the efficacy of the dynamic interference suppression strategy implemented in our method.