Classified-SLAM: A Real-Time Stereo SLAM Based on Scenario Classification and Outlier Probability Propagation for Outdoor Autonomous Vehicle

Stereo simultaneous localization and mapping (SLAM) stands as a core technology within navigation systems, commanding significant attention due to its rapid initialization and accurate, facile acquisition of semantic information. Although learning-based stereo SLAM for outdoor autonomous vehicle, where moving objects are commonly encountered, has seen significant improvements, two key challenges remain for its long-term online application: the real-time performance issues introduced by additionally semantic models, and the accurate detection of dynamic features related to pose preestimation. To address these challenges, we present classified-SLAM, a real-time stereo SLAM system based on scenario classification (SC) and outlier probability propagation, which does not rely on pose preestimation and feature depth for outlier detection. First, to accurately track dynamic features, an outlier detection method combining the affine consistency constraint with a sliding window-based probability propagation model is designed. Second, to improve the system’s real-time performance, a SC based on the optimized affine information is proposed. This is achieved by performing the semantic method combining affine information with YOLOv5s-seg model solely for highly dynamic frames classified by the SC method. Finally, to ensure the effectiveness of map for long-term autonomous applications, a map point erasure method based on outlier observation information is constructed. The extensive experiments conducted on real-world scenarios under various conditions, as well as on the KITTI odometry sequences, demonstrate that our solution outperforms state-of-the-art (SOTA) algorithms in most cases and operates in real-time during global tracking.