Egomotion estimation by fusing events and depth
Yükleniyor...
Dosyalar
Tarih
2023
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
Türk-Alman Üniversitesi Fen Bilimler Enstitüsü
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Bir robotun kendini konumlandırması, ¸cevresini algılamasıyla ba¸slar. Bir robo tun konumundan bilgi ¸cıkarmak, haritalar olu¸sturmaya, robotun genel konum de?gi¸sikli?gini hesaplamaya yardımcı olur. G¨orsel sens¨orler, d¨unyayı algılarken di?ger sens¨orler i¸cerisinde en anlamlı bilgiyi sa?glarlar. Ancak bu sens¨orler, hızla de?gi¸sen bir arka plana sahip bir depolama alanında kendini konumlandıran otonom bir robot veya bir t¨unelden ¸cıkan otonom bir ara¸c gibi zorlu senary olardan anlamlı bilgiler ¸cıkarmakta zorlanırlar. Bu ¸calı¸sma, otonom ara¸clara, alternatif sens¨orlere (¨orn., LIDAR , GPS) ihtiya¸c duymadan zorlu senaryolarda ¸calı¸smayı iyile¸stirecek ve b¨oylece maliyetleri d¨u¸s¨urecek bir algılama sistemi kazandırmaktadır. Ayrıca, bu ¸calı¸sma insan-robot etkile¸simindeki yaralanma ve kayıpları azaltmayı ama¸clamaktadır. Olay kameraları, yukarıda belirtilen senaryolarda geleneksel kameralara kıyasla ¸cok ¨onemli avantajlar sunar. Olay kameraları, do?gadan ilham alan, insan g¨oz¨un¨u taklit eden sens¨orlerdir. Bu sens¨orler, robotik algı i¸cin ola?gandı¸sı olmasa da, konumlandırma ve ego-hareket tahmini gibi durumlarda olay kameralarının benzersiz ¸cıktısını i¸slemeye y¨onelik teknikler hen¨uz olgunla¸smamı¸stır. Spesifik olarak, bu tez, bilgisayarla g¨orme tekniklerine dayalı poz tahmini ger¸cekle¸stirmek i¸cin bir olay kamerasından ve bir derinlik kamerasından alınan verileri birle¸stirmeyi ¨onermektedir. Ama¸c, otonom bir aracının algılama sistemini olu¸sturmak i¸cin olay ve derinlik kameralarının tamamlayıcı ¨ozelliklerinden yararlanmaktır. Olay kameraları y¨uksek hızlı tepki sa?glarken, derinlik kameraları sahnenin 3B yapısının tahminini basitle¸stirir. De rinlik kamerası, olay kamerasının bir sahnenin 3B temsilini toplamak i¸cin ¨u¸c¨unc¨u boyutu sa?glamasına yardımcı olur, ardından bu 3B koordinatlar poz tahmini i¸cin kullanılabilir. Ama¸c, hareket planlama ve kontrol i¸slemlerinin daha sonra g¨uvenilir bir ¸sekilde ger¸cekle¸stirilebilmesi i¸cin sa?glam bir egomotion sistemi sa?glamaktır. Onerilen yakla¸sımın ¨ozelliklerinin ve performansının, problemin ¨ anla¸sılmasını ve ¸c¨oz¨um¨un¨u ilerletmek i¸cin di?ger ¸calı¸smalarla kar¸sıla¸stırılması ama¸clanmaktadır. Ba¸slangı¸cta Nokta Bulutlarını e¸sle¸stirmek i¸cin kullanılan, iyi bilinen bir y¨ontem olan Yinelemeli En Yakın Noktalar kullanılmı¸stır. Nokta bulutu e¸sle¸stirmesinin sonu¸cları, ardı¸sık iki nokta bulutu arasındaki, onları bir birine e¸sleyebilecek transformasyonlardır. Bu nokta bulutu e¸sle¸stirmesinin ¸cıktısı, kameranın egomotion tahmini yolunu verebilir. Bahsedilen y¨ontemin avantaj ve eksikliklerini belirlemek yoluyla bu ¸calı¸sma bu alanda gelecekteki geli¸smelere yol g¨osterecektir.
For a robot to localize itself starts with perceiving its surroundings. Extracting information from the location of a robot helps to create maps, calculate the robot’s overall change in position. Visual sensors provide the most meaning ful information among other sensors when perceiving the world However, these sensors struggle to extract meaningful information from challenging scenarios, such as an autonomous robot localizing itself in a storage area with a rapidly changing background or an autonomous vehicle exiting a tunnel. This study en dows autonomous vehicles with a perception system that will improve operation in challenging scenarios without the need for alternative sensors,(e.g., LIDAR , GPS), thus decreasing costs. Furthermore, this study aims to reduce the in juries and casualties in human-robot interaction. Event cameras offer crucial advantages compared to traditional cameras in the above-mentioned scenarios. Event cameras are sensors that are inspired by nature, that mimic the human eye. Although these sensors are not unusual to robotic perception, techniques for processing the unique output of event cameras in situations such as local ization and egomotion estimation are still non-mature. Specifically, this thesis proposes to fuse data from an event camera and a depth camera to perform pose estimation based on computer vision techniques. The aim is to leverage the com plementary characteristics of event and depth cameras to build the perception system of an autonomous agent. Event cameras provide high-speed response while depth cameras simplify the estimation of the 3D structure of the scene. The depth camera assists the event camera in providing the third dimension to gather a 3D representation of a scene thereafter these 3D coordinates can be used for pose estimation. The goal is to provide a robust egomotion sys tem so that motion planning and control operations can be reliably performed afterwards. It is intended to compare the characteristics and performance of the proposed approach with other works to advance the understanding of the problem and its solution. A well-known method Iterative Closest Points, that was originally used to match the Point Clouds, has been used. The results of the point cloud matching are transformations, that will match them to each other, between two consecutive point clouds. The output of this point cloud matching can give the path thus, egomotion estimation of the camera. By identifying the advantages and shortcomings via mentioned method this study will guide future developments in this field
For a robot to localize itself starts with perceiving its surroundings. Extracting information from the location of a robot helps to create maps, calculate the robot’s overall change in position. Visual sensors provide the most meaning ful information among other sensors when perceiving the world However, these sensors struggle to extract meaningful information from challenging scenarios, such as an autonomous robot localizing itself in a storage area with a rapidly changing background or an autonomous vehicle exiting a tunnel. This study en dows autonomous vehicles with a perception system that will improve operation in challenging scenarios without the need for alternative sensors,(e.g., LIDAR , GPS), thus decreasing costs. Furthermore, this study aims to reduce the in juries and casualties in human-robot interaction. Event cameras offer crucial advantages compared to traditional cameras in the above-mentioned scenarios. Event cameras are sensors that are inspired by nature, that mimic the human eye. Although these sensors are not unusual to robotic perception, techniques for processing the unique output of event cameras in situations such as local ization and egomotion estimation are still non-mature. Specifically, this thesis proposes to fuse data from an event camera and a depth camera to perform pose estimation based on computer vision techniques. The aim is to leverage the com plementary characteristics of event and depth cameras to build the perception system of an autonomous agent. Event cameras provide high-speed response while depth cameras simplify the estimation of the 3D structure of the scene. The depth camera assists the event camera in providing the third dimension to gather a 3D representation of a scene thereafter these 3D coordinates can be used for pose estimation. The goal is to provide a robust egomotion sys tem so that motion planning and control operations can be reliably performed afterwards. It is intended to compare the characteristics and performance of the proposed approach with other works to advance the understanding of the problem and its solution. A well-known method Iterative Closest Points, that was originally used to match the Point Clouds, has been used. The results of the point cloud matching are transformations, that will match them to each other, between two consecutive point clouds. The output of this point cloud matching can give the path thus, egomotion estimation of the camera. By identifying the advantages and shortcomings via mentioned method this study will guide future developments in this field
Açıklama
Anahtar Kelimeler
Event camera, Sensor fusion, Visual odometry, Egomotion estimation, 3D computer vision, Olay kamerası, Sensör füzyonu, Görsel odometri, Ego-hareket tahmini, 3B bilgisayarla görü
Kaynak
WoS Q Değeri
Scopus Q Değeri
Cilt
Sayı
Künye
Subutay, Ebru (2023). Egomotion estimation by fusing events and depth. Türk-Alman Üniversitesi, Fen Bilimler Enstitüsü, İstanbul.
