A Semantic SLAM System Based on Object Detection and Motion Detection

In recent years, deep learning models have provided new avenues for Visual SLAM, while simultaneously introducing novel challenges, such as model missed detections and over-reliance on prior information. Therefore, to address the new issues arising subsequent to the integration of object detection models into SLAM systems, this paper proposes a novel semantic SLAM method. Specifically, this method utilizes multi-view geometry techniques to achieve compensation for missed detections of dynamic objects, thereby reducing the missed detection rate of the object detection model and enhancing the stability of front-end semantic extraction. Concurrently, it employs optical flow tracking to monitor the motion of static objects, mitigating the adverse impact on the system caused by their movement. Furthermore, a region growing algorithm is utilized to perform foreground segmentation within bounding boxes, enabling the system to maximally preserve environmental information during dense map construction. In experimental evaluations, the proposed method was tested using the TUM dataset. The results demonstrate that the proposed approach improves the system's localization accuracy and exhibits robust performance. Moreover, comparisons against state-of-the-art methods reveal its superior advantages.