Online environment segmentation based on spectral mapping
- Autores: Ricardo Vázquez Martín
- Directores de la Tesis: Antonio Bandera (dir. tes.)
- Lectura: En la Universidad de Málaga ( España ) en 2009
- Idioma: español
- Tribunal Calificador de la Tesis: Francisco Sandoval Hernández (presid.), Javier González Jiménez (secret.), Joao Fernando Cardoso Silva Sequeira (voc.), Rubén Martínez Cantín (voc.), Pablo Bustos García de Castro (voc.)
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
This Thesis addresses different issues in the autonomous navigation framework. In this aspect, the integration of different modules is necessary to fulfil this fundamental ability. Perception, localization and mapping are essential components in a navigation system. In this work, a feature-based representation is used to describe the environment, and two feature extraction algorithms have been presented for laser rangefinder and vision. Regarding the localization and mapping problem, the extracted laser features have been used to implement a Simultaneous Localization And Mapping (SLAM) algorithm. Most of the SLAM approaches relies on metric information in the mapping process, leading to some limitations related to computational complexity and accumulation of errors. In this Thesis, a new approach for map partitioning according to the environment structure is presented.
In order to interact with the environment, robots usually carry onboard sensors to collect information (e.g., sonar arrays, laser rangefinder, vision). The understanding of this sensory information is critical for other high-level tasks, such as localization and mapping. An algorithm for feature extraction and characterization from laser scan data is presented. It is based on a curvature-based method where the raw data is segmented in sets of range readings using an adaptive curvature estimation. The proposed approach presents a curvature estimator to characterize the scan contour in order to achieve a affine-invariant laser scan segmentation. On the other hand, an affine region detector based on a perceptual grouping algorithm to extract visual features has been also proposed. In this approach, the image is pre-segmented in order to obtain an initial set of colored blobs. Later, these blobs are grouped in order to simplify the image partition. Both approaches have been tested in different environments, using different laser sensors and cameras and compared to some state-of-the-art algorithms.
When no independent source of location information is available (such as Global Positioning System (GPS)) relative observations of the environment must be used. The SLAM problem solves this situation when a vehicle is moving in an unknown environment. An Extended Kalman Filter (EKF) based SLAM algorithm has been implemented using the extracted laser features. In order to solve the computational problems and the accumulation of errors when dealing with metric maps, a map partitioning algorithm has been proposed. This approach is based on appearance information achieving a map partition according to the structure of the environment. Observations gathered during the mapping process are used to build an auxiliary graph where the main property is the locality of features. The proposed map partioning algorithm employs spectral clustering to find balanced partitions in this auxiliary graph. The proposed approach provides an online submap generation to find partitions in those areas where the environment share a minimun amount of information. Finally, this algorithm is able to be used for any sort of features and sensors.
