A multimodal loop closure fusion for autonomous vehicles SLAM

Place recognition and loop closure detection are critical steps in the process of Simultaneous Localization and Mapping (SLAM). Indeed, the ability to determine whether an Autonomous Ground Vehicle (AGV) has returned to a previously visited place is highly important in the context of building a reliable SLAM system. In order to build a consistent global map and to localize the AGV with high confidence in an unknown environment, it is crucial to reduce the cumulative error generated by pose estimation. Although multiple approaches using various data sources have been proposed in order to provide an accurate pose estimation, fewer studies have focused on the integration of a multimodal process to detect loop closure. In this work, we present a novel approach to leverage multiple modalities for a robust and reliable loop closure detection. Our method is based on Similarity-Guided Particle Filtering (SGPF) for the search and validation of Loop Closure Candidates (LCCs). We validate the proposed Multimodal Loop Closure (MMLC) by using two perception modalities based on Bag-of-Words and Scan Context techniques for camera-based and LiDAR-based place recognition, respectively. The efficiency of our method has been evaluated on both KITTI and a self-collected dataset. Compared to the classical loop closure used in ORB-SLAM2, the suggested approach reduces the Absolute Trajectory Error (ATE) by up to 54% and the cumulative error during run-time by up to 62.63%. Finally, 100% of the loops are accurately detected and the ground truth distance between the current pose and the LC is less than 3 m in 98% of the cases.