Visual obstacle detection and avoidance for indoor mobile robots

Abstract

Bu calsmadaki amac kucuk boyutlu model bir arabann otonom hareket edebilenbir robot haline getirilmesidir. Bu donusumu saglamak icin birbirine bagl veonemli asamalar vardr. Bu asamalarn ilki model arabann uzerine gerekli ekipmann yerlestirilmesi ksmdr ve bu ksm Isk Universitesi Robotik ve OtonomAraclar Laboratuvar arastrmaclar tarafndan birlikte gerceklestirilmistir.Arac donanmsal olarak hazrlandktan sonra robot kontrolu ve robotla iletisimkurma amacyla gerekli programlarn kodlar yazlmstr. Tez calsmas srasnda,yazar ana program ayn anda engel tanma, engelden saknma ve sensor bilgilerinialmak amacyla cok iplikli bir yapda tasarlamstr.Bu tez calsmasnda ele alnan en onemli konu normal bir kamera kullanarakengelleri robotun hareket ettigi zeminden ayrt edebilmektir. Gelistirilen engeltanma algoritmas, ViBe (VIsual Background Extractor) [1] isimli hareketalglama algoritmasndan uyarlanmstr. Gelistirilen algoritma engelleri ayrt edebilmekicin hareket edilen zeminin modelini kullanmaktadr. Zemini modellemekamacyla her piksel lokasyonu icin bir zemin modeli saklanmaktadr. Gelistirilenalgoritma ortamdaki sk duzeyi farkllklarna, golgelere ve zemindeki goruntudegisikliklerine kars gurbuzdur. Bunun yan sra, soz konusu algoritma engelleriayrt etmek icin zemin modelini kullanan baska bir algoritmayla [2] elleisaretlenmis datalar vastasyla karslastrlmstr. Yaplan detayl deneyler sonucundaortaya ckarlan algoritmann stabil oldugu, degisik kosullarda calsabildigive diger algoritmadan daha ustun oldugu gozlenmistir.Engel tanmann yannda, tez calsmas srasnda robot icin kapal alanda engeldensaknma algoritmalar da gelistirilmistir. Yaplan deneyler robotun ses-otesi sensorler kullanarak veya sadece normal bir kameradan elde ettigi goruntuyuisleyerek kapal alanda dolasabildigini gostermistir.

This study is a part of a joint team eort to transform a small-scale model carinto an autonomous moving robot. This transformation includes several routinesthat are essential and attached together. Integration of various equipment on themodel car is the rst step of that routine which is shared among dierent thesisstudies conducted at RAVLAB (Robotics and Autonomous Vehicles Laboratory)of Isk University. Hence, the resulting hardware system which is explained inthis thesis is mostly the co-produce of RAVLAB team.The integration is followed by implementing the software required to establishcontrol and communication links between dierent units. During this thesis, theauthor has developed a multi-threaded main control software to facilitate obstacleavoiding movements of the robot while reading and analyzing sensory inputs.This thesis study mainly focuses on detection of the obstacles with visual informationcollected from an ordinary color camera. The main obstacle detectionalgorithm that is proposed in this thesis is adapted from a powerful backgroundsubtraction algorithm ViBe [1] (Visual Background Extractor). The proposedalgorithm uses the model of the ground plane in order to detect obstacles. Adierent ground plane model is kept for each pixel location as in ViBe. Theproposed algorithm is robust against illumination dierences, shadows and thechanges in the appearance of the ground plane. A comparison is provided (usingground truth data) with another obstacle detection algorithm [2] which also usesa ground plane based model to detect obstacles. The results of the proposedalgorithm under dierent conditions compared to a counterpart.In addition to the obstacle detection, during this study two obstacle avoidancealgorithms are developed to facilitate navigation of the robot in indoor environments.The experiments show that the robot is able to move while avoidingobstacles by using ultrasonic sensors as well as using the visual camera input.