LTD-SLAM: Multi-Scale Ligh-Weight Model and Adaptive Thresholds Based Dynamic Object Detection for Semantic Visual SLAM
Lang Qian, Jianmei Su, Ling Wu · 2024
The cumulative error of pose estimation for intelligent mobile robot in dynamic environment will result in low positioning accuracy. Most of the existing methods that incorporate neural networks to solve dynamic problems require high computational support and predefined types of dynamic objects. LTD-SLAM is a lightweight and real-time semantic visual SLAM system based on the ORB-SLAM3 framework. In LTD-SLAM, a lightweight object detection network is designed to provide semantic information and improve the real-time performance of SLAM. Then an adaptive method is proposed for distinguishing background and foreground points in dynamic environments without requiring predefined dynamic objects, utilizing local optical flow to eliminate dynamic points. Finally, the average reprojection error of foreground points across multiple frames is calculated, and a threshold is obtained by combining with the probability density function, thereby enabling the recovery of dynamic points in low-dynamic scenarios. The experimental results show that compared with other dynamic SLAM methods, the proposed method can obtain almost the best results in high and low dynamic scenes with higher real-time performance.