DRV-SLAM: An Adaptive Real-Time Semantic Visual SLAM Based on Instance Segmentation Toward Dynamic Environments

Traditional simultaneous localization and mapping (SLAM) methodologies predominantly rely on the assumption of a static environment. This constraint limits the applicability of most visual SLAM systems in various real-world scenarios. In this paper, we introduce a real-time semantic visual SLAM algorithm tailored for complex dynamic environments (DRV-SLAM). DRV-SLAM leverages image analysis to identify potential moving objects and determine their current motion states. By dynamically adjusting the rejection of unreliable dynamic feature points based on the proportion of potential moving objects in the environment, DRV-SLAM significantly enhances the system's localization accuracy and robustness in complex dynamic environments. Additionally, DRV-SLAM employs a dense mapping approach that combines global downsampling and targeted object data enhancement. This method effectively reduces the memory footprint of dense point cloud maps, enabling DRV-SLAM to efficiently construct large-scale dense point cloud maps in diverse scenarios. Experimental results show that in a highly dynamic environment, the DRV-SLAM algorithm shows an order of magnitude performance improvement compared to the traditional ORB-SLAM2 algorithm. The performance index of absolute trajectory error is significantly improved by more than 98%, and DRV-SLAM is currently one of the most real-time, accurate and robust systems for dynamic scenes.