Publicação
Application of semantic segmentation through data acquired from sensors
| Resumo: | Today, AI is very important in our lives as its used all around us without our knowledge. From simple things such as personal assistants like Alexa and Siri, and advertising algorithms focusing on our tastes - Netflix on the recommendation of movies or, even more common, the presentation of advertising based on our search history -, to robots and to smart houses, cities or even vehicles. The presence of AI is increasing and even if we are still far away from our ’General AI’ ideology, a machine capable of anything autonomously, each day we get closer. In the last decade multiple applications of AI have been through breakthroughs. For example, the first implementations of autonomous vehicles were introduced by Tesla and other companies. A number of discoveries must have been made to achieve this revolution of AI performance and, among them, is two of the most important developments: Object Detection and Semantic Segmentation, closely related to each other. These are responsible for understanding the environment so the machine can take actions, being the latter an improvement of the first in terms of sensibility error associated to each entity detected as well as being able to detect its corresponding type, in a pixel level. These machines require more and more data to analyse, having many types of sensors in order to collect information, such as radars, cameras, LiDAR, among others. This work falls in the study of the use of Semantic Segmentation techniques and its application on categorising data from image related sensors in order to explain its breakthroughs and challenges, as well as improving and overcoming such obstacles. Data will consist mainly of scans from outdoor/self-driving cars POV (KITTI360) with the ability to be used with other types of data such as indoor scans (COCO), to explain both road and more day-to-day images semantic compositions, applied on a state-of-art solution. Consecutively we will perform a process of optimisation in order to reduce computation costs. Currently the works of DeepLab (with the research of deeplabv3[1]) have achieved a high success on Semantic Segmentation overcoming previous problems such as handling component boundaries with more refined lines while keeping it fairly easy to run on more less powerful machines, being the start point for this work. |
|---|---|
| Autores principais: | Monteiro, Filipe Pimenta Oliveira |
| Assunto: | Image segmentation Semantic segmentation Deep learning Self-driving Automotive security Segmentação semântica de imagens Veículos autónomos Redes neuronais convolucionais Segurança automóvel |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Today, AI is very important in our lives as its used all around us without our knowledge. From simple things such as personal assistants like Alexa and Siri, and advertising algorithms focusing on our tastes - Netflix on the recommendation of movies or, even more common, the presentation of advertising based on our search history -, to robots and to smart houses, cities or even vehicles. The presence of AI is increasing and even if we are still far away from our ’General AI’ ideology, a machine capable of anything autonomously, each day we get closer. In the last decade multiple applications of AI have been through breakthroughs. For example, the first implementations of autonomous vehicles were introduced by Tesla and other companies. A number of discoveries must have been made to achieve this revolution of AI performance and, among them, is two of the most important developments: Object Detection and Semantic Segmentation, closely related to each other. These are responsible for understanding the environment so the machine can take actions, being the latter an improvement of the first in terms of sensibility error associated to each entity detected as well as being able to detect its corresponding type, in a pixel level. These machines require more and more data to analyse, having many types of sensors in order to collect information, such as radars, cameras, LiDAR, among others. This work falls in the study of the use of Semantic Segmentation techniques and its application on categorising data from image related sensors in order to explain its breakthroughs and challenges, as well as improving and overcoming such obstacles. Data will consist mainly of scans from outdoor/self-driving cars POV (KITTI360) with the ability to be used with other types of data such as indoor scans (COCO), to explain both road and more day-to-day images semantic compositions, applied on a state-of-art solution. Consecutively we will perform a process of optimisation in order to reduce computation costs. Currently the works of DeepLab (with the research of deeplabv3[1]) have achieved a high success on Semantic Segmentation overcoming previous problems such as handling component boundaries with more refined lines while keeping it fairly easy to run on more less powerful machines, being the start point for this work. |
|---|