Scholarlink

<p>Visual simultaneous localization and mapping (V-SLAM) algorithms usually assume a static working environment. However, in reality, there are always moving objects that can significantly affect the accuracy of the V-SLAM algorithm and even cause it to fail. In practical applications, the V-SLAM system may encounter numerous moving objects, such as vehicles in autonomous driving scenarios and pedestrians in augmented reality (AR) applications. It is essential to accurately recognizing the real state of dynamic objects, whether they are moving or static, to ensure the proper operation of the V-SLAM system in high-dynamic scenes. This paper proposes DOA-SLAM, a method that reduces the impact of dynamic objects on V-SLAM with minimal time cost. DOA-SLAM initially employs real-time instance segmentation to detect potential moving objects in the scene. Key points are then filtered based on segmentation masks to ensure that sufficient static feature points are retained for tracking. Subsequently, a lightweight object association method is proposed to identify the same object in different frames. Finally, the method identifies objects that have moved between frames through a motion consistency check. The proposed DOA-SLAM method excludes dynamic objects from pose estimation and does not add their observation information to the map. Experiments were conducted on the KITTI Odometry and TartanAir Shibuya datasets. The results demonstrate that the method improves the accuracy of pose estimation for V-SLAM in dynamic scenes and creates a more uniform representation of the geometric information in the environment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DOA-SLAM: An Efficient Stereo Visual SLAM System in Dynamic Environment

  • 期刊论文

  • 发表日期：

    2025年4月12日

查看全文
  • Zhaoqian Jia,
  • Yixiao Ma,
  • Junwen Lai,
  • Zhiguo Wang

摘要

Visual simultaneous localization and mapping (V-SLAM) algorithms usually assume a static working environment. However, in reality, there are always moving objects that can significantly affect the accuracy of the V-SLAM algorithm and even cause it to fail. In practical applications, the V-SLAM system may encounter numerous moving objects, such as vehicles in autonomous driving scenarios and pedestrians in augmented reality (AR) applications. It is essential to accurately recognizing the real state of dynamic objects, whether they are moving or static, to ensure the proper operation of the V-SLAM system in high-dynamic scenes. This paper proposes DOA-SLAM, a method that reduces the impact of dynamic objects on V-SLAM with minimal time cost. DOA-SLAM initially employs real-time instance segmentation to detect potential moving objects in the scene. Key points are then filtered based on segmentation masks to ensure that sufficient static feature points are retained for tracking. Subsequently, a lightweight object association method is proposed to identify the same object in different frames. Finally, the method identifies objects that have moved between frames through a motion consistency check. The proposed DOA-SLAM method excludes dynamic objects from pose estimation and does not add their observation information to the map. Experiments were conducted on the KITTI Odometry and TartanAir Shibuya datasets. The results demonstrate that the method improves the accuracy of pose estimation for V-SLAM in dynamic scenes and creates a more uniform representation of the geometric information in the environment.