An instance-aware segmentation and optical flow based DVI-SLAM for dynamic environments

Simultaneous Localisation and Mapping (SLAM) is a fundamental building block for markerless Augmented Reality systems. However, most conventional SLAM systems exhibit poor performance in real-time environments because of the influence of dynamic objects in unstructured surroundings. We are proposing a new Dynamic Visual Inertial SLAM system called DVI-SLAM to address the challenges of dynamic content presented in complex urban scenes. Leveraging instance-aware segmentation and optical flow methods, our system can detect and exclude actively moving objects from the tracking process. In addition to excluding features of moving objects from the pose estimation process, the system also detects, tracks, and generates static-map of the surrounding environment. This helps develop a robust AR system or mobile robot system to effectively handle complex outdoor urban scene scenarios. The proposed DVI-SLAM framework was tested on the TUM RGB-D and TUM-VI datasets as well as in real-world environments. The results showed an improvement in ATE’s Root Mean Square Error (RMSE) up to 24.80 and 32.34% compared with ORB-SLAM3, on monocular-inertial and stereo-inertial sequences of the TUM-VI dataset, respectively. Similarly, an improvement up to 64.89 and 73.75% was achieved for real-time monocular and RGB-D inertial sequences of real-time environments. Further, the sequence with a lot of dynamic content, such as the TUM-RGBD dynamic dataset, showed up to 99.4% improvement in ATE RMSE. The results demonstrated that DVI-SLAM significantly enhances Absolute Trajectory Error accuracy as compared to the state-of-the-art ORB-SLAM3 method. Extensive experiments evaluation shows that DVI-SLAM performs robustly across both static and dynamic scenarios, while maintaining competitive performance in static and low-dynamic scenes, it delivers significant improvements in highly dynamic environments.