Sensor Fusion SLAM: An Efficient and Robust SLAM system for Dynamic Environments

In response to challenges such as variable lighting conditions, sparse texture information, and dynamic, unstructured environments, this paper introduces an innovative radarvision-inertial odometry system for Simultaneous Localization and Mapping (SLAM) in complex scenarios. The system integrates LiDAR, vision, and inertial sensors, capitalizing on their unique attributes to deliver high-frequency, precise pose estimation and map construction. A key component is the dynamic point cloud filtering module, which employs a fully convolutional neural network to detect, segment, and mitigate the influence of dynamic point clouds on feature matching and mapping processes. For the LiDAR inertial odometry, we propose a novel approach to point cloud feature extraction based on surface concepts, coupled with adaptive filtering and an adaptive residual algorithm. These enhancements not only improve the quality of point clouds and mapping accuracy but also conserve computational resources and bolster system robustness. The vision inertial odometry component benefits from optimized sliding window computations, which enhance the system’s computational efficiency and numerical stability. Extensive testing on the EUROC and M2DGR datasets, as well as trials using mobile robotic platforms in real-world campus settings, has demonstrated that our multi-sensor fusion system outperforms conventional SLAM systems in accuracy and robustness. These results underscore the system’s potential for navigating and mapping in challenging environments.