Fast LiDAR data registration using GPUs
No Thumbnail Available
Date
2023-10-18
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Pontificia Universidad Católica del Perú
Abstract
En los últimos años, la llegada de las cámaras de profundidad de bajo costo y sensores LiDAR ha incentivado a las industrias a invertir en estas tecnologías, lo cual incluye también mayor interés en investigaciones sobre procesamiento digital de señales. En esta ocasión, la reconstrucción tridimensional de túneles mineros utilizando LiDARs y un robot de auto-navegación ha sido propuesta como proyecto de investigación, y el presente trabajo forma parte en cargándose del alineamiento de nubes de puntos tridimensionales en tiempo real, un proceso que es más conocido como Registro de Nubes de Puntos. Existen muchos algoritmos que pueden resolver este problema, pero para el proyecto, el algoritmo solo necesita calcular la alineación fina y rígida. Al comparar los algoritmos de registro más avanzados, se encontró que el popular algoritmo ICP es el más adecuado para este caso debido a su alta robustez y eficiencia. Dentro de este algoritmo, se encuentran 3 pasos simples: relación, minimización y transformación, junto con una colección de variaciones de estos pasos que han sido desarrolladas a lo largo de las últimas décadas. Basándose en esto, en este trabajo se diseñó e implementó un algoritmo ICP paralelo en CPU y GPU. Además, las optimizaciones en recursos de memoria, ocupación de núcleos y el uso de la técnica de desenrollado de bucles para la implementación en GPU permiten que la implementación propuesta deI ICP alcance un rendimiento 95 veces más rápido que implementaciones de CPU altamente optimizadas.
In the last years, the advent of low-cost depth cameras and LiDAR sensors has encouraged industries to invest in these technologies, which includes research about digital signal processing. This time, 3D surfaces reconstructions of mining tunnels using LiDARs and a self-navigation robot have been proposed as a research project and this work is in charge of performing the alignment of data that come from the LiDARs in real-time, a process that is most known as 3D Point Cloud Registration. Many algorithms can solve this problem, but for the project, the algorithm only needs to compute for fine-rigid alignment. Comparing state-of-the-art registration algorithms, it is found that the popular ICP algorithm is the best suited for this case since its high robustness and efficiency. Inside this algorithm, there are 3 simple steps: matching, minimization, and transformation, and a diverse collection of variations of these steps that have been developed through the last decades. Based on this, a parallel ICP algorithm is designed and implemented in CPU and GPU. Moreover, optimizations in memory resources, cores occupancy, and the usage of the loop unrolling technique for the GPU implementation lead the proposed ICP implementation to reach a performance of 95 times faster than highly-optimized CPU implementations.
In the last years, the advent of low-cost depth cameras and LiDAR sensors has encouraged industries to invest in these technologies, which includes research about digital signal processing. This time, 3D surfaces reconstructions of mining tunnels using LiDARs and a self-navigation robot have been proposed as a research project and this work is in charge of performing the alignment of data that come from the LiDARs in real-time, a process that is most known as 3D Point Cloud Registration. Many algorithms can solve this problem, but for the project, the algorithm only needs to compute for fine-rigid alignment. Comparing state-of-the-art registration algorithms, it is found that the popular ICP algorithm is the best suited for this case since its high robustness and efficiency. Inside this algorithm, there are 3 simple steps: matching, minimization, and transformation, and a diverse collection of variations of these steps that have been developed through the last decades. Based on this, a parallel ICP algorithm is designed and implemented in CPU and GPU. Moreover, optimizations in memory resources, cores occupancy, and the usage of the loop unrolling technique for the GPU implementation lead the proposed ICP implementation to reach a performance of 95 times faster than highly-optimized CPU implementations.
Description
Keywords
Sensores inteligentes, Algoritmos--Aplicaciones, Industria minera
Citation
Collections
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwised noted, this item's license is described as info:eu-repo/semantics/openAccess