Lee, Joseph Sung (2016-05). Appearance and Geometry Assisted Visual Navigation in Urban Areas. Doctoral Dissertation. Thesis uri icon

abstract

  • Navigation is a fundamental task for mobile robots in applications such as exploration, surveillance, and search and rescue. The task involves solving the simultaneous localization and mapping (SLAM) problem, where a map of the environment is constructed. In order for this map to be useful for a given application, a suitable scene representation needs to be defined that allows spatial information sharing between robots and also between humans and robots. High-level scene representations have the benefit of being more robust and having higher exchangeability for interpretation. With the aim of higher level scene representation, in this work we explore high-level landmarks and their usage using geometric and appearance information to assist mobile robot navigation in urban areas. In visual SLAM, image registration is a key problem. While feature-based methods such as scale-invariant feature transform (SIFT) matching are popular, they do not utilize appearance information as a whole and will suffer from low-resolution images. We study appearance-based methods and propose a scale-space integrated Lucas-Kanade's method that can estimate geometric transformations and also take into account image appearance with different resolutions. We compare our method against state-of-the-art methods and show that our method can register images efficiently with high accuracy. In urban areas, planar building facades (PBFs) are basic components of the quasirectilinear environment. Hence, segmentation and mapping of PBFs can increase a robot's abilities of scene understanding and localization. We propose a vision-based PBF segmentation and mapping technique that combines both appearance and geometric constraints to segment out planar regions. Then, geometric constraints such as reprojection errors, orientation constraints, and coplanarity constraints are used in an optimization process to improve the mapping of PBFs. A major issue in monocular visual SLAM is scale drift. While depth sensors, such as lidar, are free from scale drift, this type of sensors are usually more expensive compared to cameras. To enable low-cost mobile robots equipped with monocular cameras to obtain accurate position information, we use a 2D lidar map to rectify imprecise visual SLAM results using planar structures. We propose a two-step optimization approach assisted by a penalty function to improve on low-quality local minima results. Robot paths for navigation can be either automatically generated by a motion planning algorithm or provided by a human. In both cases, a scene representation of the environment, i.e., a map, is useful to specify meaningful tasks for the robot. However, SLAM results usually produce a sparse scene representation that consists of low-level landmarks, such as point clouds, which are neither convenient nor intuitive to use for task specification. We present a system that allows users to program mobile robots using high-level landmarks from appearance data.
  • Navigation is a fundamental task for mobile robots in applications such as exploration, surveillance, and search and rescue. The task involves solving the simultaneous localization and mapping (SLAM) problem, where a map of the environment is constructed. In order for this map to be useful for a given application, a suitable scene representation needs to be defined that allows spatial information sharing between robots and also between humans and robots. High-level scene representations have the benefit of being more robust and having higher exchangeability for interpretation. With the aim of higher level scene representation, in this work we explore high-level landmarks and their usage using geometric and appearance information to assist mobile robot navigation in urban areas.

    In visual SLAM, image registration is a key problem. While feature-based methods such as scale-invariant feature transform (SIFT) matching are popular, they do not utilize appearance information as a whole and will suffer from low-resolution images. We study appearance-based methods and propose a scale-space integrated Lucas-Kanade's method that can estimate geometric transformations and also take into account image appearance with different resolutions. We compare our method against state-of-the-art methods and show that our method can register images efficiently with high accuracy.

    In urban areas, planar building facades (PBFs) are basic components of the quasirectilinear environment. Hence, segmentation and mapping of PBFs can increase a robot's abilities of scene understanding and localization. We propose a vision-based PBF segmentation and mapping technique that combines both appearance and geometric constraints to segment out planar regions. Then, geometric constraints such as reprojection errors, orientation constraints, and coplanarity constraints are used in an optimization process to improve the mapping of PBFs.

    A major issue in monocular visual SLAM is scale drift. While depth sensors, such as lidar, are free from scale drift, this type of sensors are usually more expensive compared to cameras. To enable low-cost mobile robots equipped with monocular cameras to obtain accurate position information, we use a 2D lidar map to rectify imprecise visual SLAM results using planar structures. We propose a two-step optimization approach assisted by a penalty function to improve on low-quality local minima results.

    Robot paths for navigation can be either automatically generated by a motion planning algorithm or provided by a human. In both cases, a scene representation of the environment, i.e., a map, is useful to specify meaningful tasks for the robot. However, SLAM results usually produce a sparse scene representation that consists of low-level landmarks, such as point clouds, which are neither convenient nor intuitive to use for task specification. We present a system that allows users to program mobile robots using high-level landmarks from appearance data.

publication date

  • May 2016