Publication: Técnicas de Machine Learning para conducción
Loading...
Official URL
Full text at PDC
Publication Date
2019
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este trabajo tiene como objetivo principal trabajar con un tema tan de moda como la inteligencia artificial aplicada a la conducción autónoma. Este campo es bastante amplio y existen multitud de problemas que se pueden resolver a través de la inteligencia artificial. En nuestro caso nos hemos centrado concretamente en el problema del cálculo de la profundidad de los elementos existentes en un campo visual. Para resolver dicho problema existen multitud de arquitecturas de redes neuronales convolucionales, todas ellas basadas en el procesamiento de un par de imágenes en estéreo. De este procesamiento se obtiene un mapa de disparidad que se puede traducir en una imagen que representa la distancia existente del coche a cada uno de los píxeles que la representa.
Además, hemos tenido como objetivo utilizar OpenVINO, una herramienta que acelera la velocidad con la que se puede realizar inferencia sobre una red neuronal. Esto implica que todos los problemas que requieran de inteligencia artificial puedan tener una respuesta lo más rápida posible, uno de los puntos más críticos en este campo. Por otro lado, utilizar esta tecnología junto a elementos de cómputo vectorial de bajo consumo energético, como puede ser Movidius, puede ser una buena combinación para solventar el problema del consumo de energía por los elementos de cómputo hardware a bordo del vehículo.
Teniendo en cuenta estos objetivos, nuestro trabajo primeramente trata sobre el estado del arte en cuanto a inteligencia artificial y sus subramas, conducción autónoma y tecnologías software que se utilizan para afrontar estos temas. Posteriormente presentamos el trabajo previo con redes neuronales de clasificación como primer acercamiento a nuestro objetivo principal. A continuación se expone el trabajo realizado con redes neuronales para cálculo de disparidad, sus resultados y uso con OpenVINO. Finalmente se realiza una conclusión acerca de los resultados obtenidos.
The main objective of this work is to work with a topic as fashionable as artificial intelligence applied to autonomous driving. This field is quite broad and there are many problems that can be solved through artificial intelligence. In our case, we have focused specifically on the problem of calculating the depth of the existing elements in a visual field. To solve this problem there are a multitude of convolutional neural network architectures, all based on the processing of a pair of images in stereo. From this processing a disparity map is obtained that can be translated into an image that represents the existing distance of the car to each of the pixels that represents it. In addition, we have aimed to use OpenVINO, a tool that accelerates the speed with which inference can be made over a neural network. This implies that all the problems that require artificial intelligence can have an answer as quickly as possible, one of the most critical points in this field. On the other hand, using this technology together with elements of vector computing of low energy consumption, such as Movidius, can be a good combination to solve the problem of energy consumption by the elements of hardware computation on board the vehicle. In consideration of these objectives, our work will first talk about the state of the art in terms of artificial intelligence and its sub-branches, autonomous driving and software technologies that are used to face these issues. Later we present the previous work with neural networks of classification as the first approach to our main objective. Next, the work done with neural networks for disparity calculation, its results and use with OpenVINO is exposed and finally a final conclusion about the obtained results is made.
The main objective of this work is to work with a topic as fashionable as artificial intelligence applied to autonomous driving. This field is quite broad and there are many problems that can be solved through artificial intelligence. In our case, we have focused specifically on the problem of calculating the depth of the existing elements in a visual field. To solve this problem there are a multitude of convolutional neural network architectures, all based on the processing of a pair of images in stereo. From this processing a disparity map is obtained that can be translated into an image that represents the existing distance of the car to each of the pixels that represents it. In addition, we have aimed to use OpenVINO, a tool that accelerates the speed with which inference can be made over a neural network. This implies that all the problems that require artificial intelligence can have an answer as quickly as possible, one of the most critical points in this field. On the other hand, using this technology together with elements of vector computing of low energy consumption, such as Movidius, can be a good combination to solve the problem of energy consumption by the elements of hardware computation on board the vehicle. In consideration of these objectives, our work will first talk about the state of the art in terms of artificial intelligence and its sub-branches, autonomous driving and software technologies that are used to face these issues. Later we present the previous work with neural networks of classification as the first approach to our main objective. Next, the work done with neural networks for disparity calculation, its results and use with OpenVINO is exposed and finally a final conclusion about the obtained results is made.
Description
Trabajo de Fin de Grado, Universidad Complutense, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, Curso 2018/2019