Deep Learning geométrico aplicado a datos LiDAR
Fecha
2021-05-07
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Jaén: Universidad de Jaén
Resumen
La segmentación semántica de nubes de puntos se ha consolidado como un problema clásico de clasificación del
ámbito del aprendizaje automático, dando paso al geometric deep learning. Los datos LiDAR son un tipo especial
de nubes de puntos que son generadas utilizando escáneres LiDAR, capaces de captar más información del
entorno que un escáner convencional. Sin embargo, este tipo de datos continúan tratándose como nubes
convencionales, ignorándose este añadido de información.
Así, el objetivo principal de este proyecto es el de analizar las principales técnicas para tratamiento inteligente de
nubes de puntos, y estudiar la adaptabilidad de estos métodos para beneficiarse de la información adicional que
caracteriza a los datos LiDAR. Se propone una batería de experimentos para medir la ganancia de rendimiento
que se produce al utilizar este tipo de datos con respecto al uso convencional.
Point cloud semantic segmentation has become a standard classification problem in the machine learning field, motivating the emergence of geometric deep learning. LiDAR point clouds are a special kind of clouds that are generated using LiDAR scanners, sensors able to capture much more information than a traditional scanner. However, those enriched point clouds are being treated as conventionals one, ignoring some increased-value information that is scanned with LiDAR. So, the main objective of this project is to analyze the main techniques for point cloud intelligent processing, and to study the adaptability of those methods to benefit from the extra information that are characteristic from LiDAR data. A batch of experiments is defined to measure the gain in the performance when using this type of data with respect to the conventional use.
Point cloud semantic segmentation has become a standard classification problem in the machine learning field, motivating the emergence of geometric deep learning. LiDAR point clouds are a special kind of clouds that are generated using LiDAR scanners, sensors able to capture much more information than a traditional scanner. However, those enriched point clouds are being treated as conventionals one, ignoring some increased-value information that is scanned with LiDAR. So, the main objective of this project is to analyze the main techniques for point cloud intelligent processing, and to study the adaptability of those methods to benefit from the extra information that are characteristic from LiDAR data. A batch of experiments is defined to measure the gain in the performance when using this type of data with respect to the conventional use.
Descripción
Palabras clave
Tratamiento Inteligente de la Información