Publication: Segmentación semántica para la conducción autónoma mediante aprendizaje por refuerzo
Loading...
Identifiers
Publication date
2020-10
Defense date
2020-11-24
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
La Conducción Autónoma es una tecnología que está ganando cada vez más inercia. Se están haciendo avances y estudios de forma constante para el despliegue de la Conducción Autónoma en el mundo del automóvil en un futuro cercano. Se está empujando esta tecnología para que los vehículos puedan conducir con mayor grado de autonomía.
Existen diferentes técnicas que se puedes usar a la hora de crear un agente que controle un vehículo, y una de las técnicas que se están empezando a estudiar es el Aprendizaje por Refuerzo. Esta técnica podría superar al resto en la Conducción Autónoma, gracias a su capacidad para crear agentes que se automejoran y que exploran todas las posibilidades por su cuenta.
En mi investigación anterior “Deep Reinforcement Learning Aplicado a la Conducción Autónoma” se creó un agente mediante Aprendizaje por Refuerzo que era capaz de conducir en una carretera con curvas en un desierto.
En este proyecto se ha ido un paso más allá y se ha creado un agente capaz de conducir y seguir rutas en una ciudad con intersecciones. El agente se ha creado con el algoritmo de Soft Actor-Critic. La entrada del agente es una imagen codificada con un Variational Autoencoder. Sin embargo, un Autoencoder entrenado de la manera tradicional, para reconstruir la entrada original, no es capaz de afrontar esta tarea. Se ha entrenado el Autoencoder para reconstruir una versión de la entrada con Segmentación Semántica de forma que se codifique información más útil. En este proyecto se comparan ambas aproximaciones, y se comprueba si la versión con Segmentación Semántica obtiene mejores resultados. Para comparar ambos modelos se evalúan en función del tiempo medio entre fallos del agente, del porcentaje de autonomía y de la recompensa del agente (la cual depende de su velocidad, de lo centrado que vaya en la carretera y de la diferencia entre el ángulo del vehículo y la dirección de la carretera).
Autonomous Driving is a technology that is gaining more and more inertia. Progress and studies are constantly being made for the deployment of Autonomous Driving in the automotive world in the near future. This technology is being pushed so that vehicles can drive with a greater degree of autonomy. There are different techniques that can be used when creating an agent to control a vehicle, and one of the techniques that are beginning to be studied is Reinforcement Learning. This technique could outperform the rest in Autonomous Driving, thanks to its ability to create self-improving agents who explore all possibilities on their own. In my previous research “Deep Reinforcement Learning Applied to Autonomous Driving” an agent was created using Reinforcement Learning which was capable of driving on a curvy road in a desert. This project has gone a step further and created an agent capable of driving and following routes in a city with intersections. The agent has been created with the Soft Actor-Critic algorithm. The agent input is an image encoded with a Variational Autoencoder. However, an Autoencoder trained in the traditional way, trying to reconstruct the original input, is not able to perform this task. The Autoencoder has been trained to reconstruct a version of the input with Semantic Segmentation so that more useful information is encoded. In this project both approaches are compared, and it is checked if the version with Semantic Segmentation obtains better results. To compare both models, they are evaluated based on the mean time between agent failures, the percentage of autonomy and the agent's reward (which depends on its speed, how centered it is on the road and the difference between the angle of the vehicle and the direction of the road).
Autonomous Driving is a technology that is gaining more and more inertia. Progress and studies are constantly being made for the deployment of Autonomous Driving in the automotive world in the near future. This technology is being pushed so that vehicles can drive with a greater degree of autonomy. There are different techniques that can be used when creating an agent to control a vehicle, and one of the techniques that are beginning to be studied is Reinforcement Learning. This technique could outperform the rest in Autonomous Driving, thanks to its ability to create self-improving agents who explore all possibilities on their own. In my previous research “Deep Reinforcement Learning Applied to Autonomous Driving” an agent was created using Reinforcement Learning which was capable of driving on a curvy road in a desert. This project has gone a step further and created an agent capable of driving and following routes in a city with intersections. The agent has been created with the Soft Actor-Critic algorithm. The agent input is an image encoded with a Variational Autoencoder. However, an Autoencoder trained in the traditional way, trying to reconstruct the original input, is not able to perform this task. The Autoencoder has been trained to reconstruct a version of the input with Semantic Segmentation so that more useful information is encoded. In this project both approaches are compared, and it is checked if the version with Semantic Segmentation obtains better results. To compare both models, they are evaluated based on the mean time between agent failures, the percentage of autonomy and the agent's reward (which depends on its speed, how centered it is on the road and the difference between the angle of the vehicle and the direction of the road).
Description
Keywords
Aprendizaje por refuerzo profundo, CityScapes, Conducción autónoma, Segmentación semántica, Soft actor-critic, Variational autoencoder, Autonomous driving, Deep reinforcement learning, Semantic segmentation