A Lightweight Visual Odometry Based on Object Detection for Dynamic Environment

The current Simultaneous Localization and Mapping (SLAM) systems are primarily based on the assumption of scene rigidity. However, the interference of dynamic objects in real-world environments and the computational limitations of small-scale computers pose significant challenges for the application of SLAM. To address these issues, this paper proposes a dynamic visual SLAM (VSLAM) system characterized by low computational cost, good stability, and maintained localization accuracy. This system first adopts an improved lightweight YOLOv8n algorithm for rapid detection of dynamic targets. Subsequently, a mask generation algorithm is proposed, which categorizes detection results into dynamic and static regions and eliminates feature points in the dynamic area using masks generated from bounding boxes, while preserving the features of static target areas. Finally, testing results on the TUM dataset demonstrate that compared to ORB-SLAM3, our system achieves an average reduction of 90.6% in the root mean square error of absolute trajectory error in highly dynamic sequences. Compared to DynaSLAM, our system significantly reduces the average tracking time while maintaining similar localization accuracy. Deployment tests on mobile platforms further confirm the system’s lower computational demands and robust operational performance.