Localization and mapping in the real scene where movable objects exist is a challenging task, however, most SLAM approaches generally assume that the scene is static. This assumption leads to building a map with movable objects, causing the localization to fail. We propose MODS, a movable object discrimination strategy which generates a knowledge base containing different attributes of an object, classify the object according to various attributes in the knowledge base, and take different measures for different categories. By using MODS, lidar data can be used in a selective manner to avoid data redundancy and reduce the probability of localization failure. We implement a complete SLAM system based on MODS, verify the influence of movable objects on SLAM with KITTI dataset, and also estimate our SLAM system. The qualitative experimental results show that our SLAM system can effectively eliminate the influence of movable objects on mapping and localization. Moreover, the quantitative experiment results show that the time required for localization was reduced by 16.0% on average compared with the state-of-the-art, while the loop-closure can still be correctly detected.
Paper
Full text
StaticSLAM: A Dynamic Object-free Accurate SLAM System
Semantic Scholar · Engineering · 2019
Abstract
Localization and mapping in the real scene where movable objects exist is a challenging task, however, most SLAM approaches generally assume that the scene is static. This assumption leads to building a map with movable objects, causing the localization to fail. We propose MODS, a movable object discrimination strategy which generates a knowledge base containing different attributes of an object, classify the object according to various attributes in the knowledge base, and take different measures for different categories. By using MODS, lidar data can be used in a selective manner to avoid data redundancy and reduce the probability of localization failure. We implement a complete SLAM system based on MODS, verify the influence of movable objects on SLAM with KITTI dataset, and also estimate our SLAM system. The qualitative experimental results show that our SLAM system can effectively eliminate the influence of movable objects on mapping and localization. Moreover, the quantitative experiment results show that the time required for localization was reduced by 16.0% on average compared with the state-of-the-art, while the loop-closure can still be correctly detected.