Monokulaarsetelt piltidelt tee tuvastamine ja eristamine kasutades tehisnärvivõrke
Files
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Teede eristamine on oluline osa iseseisvatest navigatsioonisüsteemidest, mis aitavad robotitel ja autonoomsetel sõidukitel maapinnal liikuda. See on kasutusel erinevates seotud alamülesannetes, näiteks võimalike valiidsete liikumisteede leidmisel, takistusega kokkupõrke vältimisel ja teel asuvate objektide avastamisel.Selle töö eesmärk on uurida eksisteerivaid teede tuvastamise ja eristamise võtteid ning pakkuda välja alternatiivne lahendus selle teostamiseks.Töö jaoks loodi 5300-pildine andmestik ilma lisainfota teepiltidest. Lisaks tehti kokkuvõte juba eksisteerivatest teepiltide andmestikest. Töös pakume erinevates keskkondades asuvate teede piltide klassifitseerimiseks välja LeNet-5’l põhineva tehisnärvivõrgu. Samuti esitleme FCN-8’l põhinevat mudelit pikslipõhiseks pildituvastuseks.
Road recognition is one of the important aspects in Autonomous Navigation Systems. These systems help to navigate the autonomous vehicle and robot on the ground. Further, road detection is useful in related sub-tasks such as finding valid road path where the robot/vehicle can go, for supportive driverless vehicles, preventing the collision with the obstacle, object detection on the road, and others.The goal of this thesis is to examine existing road detection and recognition techniques and propose an alternative solution for road classification and detection task.Our contribution consists of several parts. Firstly, we released the road images dataset with approximately 5,300 unlabeled road images. Secondly, we summarized the information about the existing road images datasets. Thirdly, we proposed the convolutional LeNet-5-based neural network for the road image classification for various environments. Finally, our FCN-8-based model for pixel-wise image recognition has been presented.
Road recognition is one of the important aspects in Autonomous Navigation Systems. These systems help to navigate the autonomous vehicle and robot on the ground. Further, road detection is useful in related sub-tasks such as finding valid road path where the robot/vehicle can go, for supportive driverless vehicles, preventing the collision with the obstacle, object detection on the road, and others.The goal of this thesis is to examine existing road detection and recognition techniques and propose an alternative solution for road classification and detection task.Our contribution consists of several parts. Firstly, we released the road images dataset with approximately 5,300 unlabeled road images. Secondly, we summarized the information about the existing road images datasets. Thirdly, we proposed the convolutional LeNet-5-based neural network for the road image classification for various environments. Finally, our FCN-8-based model for pixel-wise image recognition has been presented.