Citation Number
4
Views
15
Downloands
6
Bilgisayarlı görü teknolojisi 1960’lı yıllardan itibaren gelişmeye başladı ve günümüzde bu alanda oldukça ilerleme katedildi. Bugün gerçek zamanlı görüntüler üzerinde arabalar, insanlar gibi farklı hareket halinde olan nesneler takip edilebilmektedir. Fakat doğrusal olmayan yörüngelerde ve çok hızlı hareket eden küçük cisimlerin tespiti ve takibi çoğu durum için daha doğrusal, normal hızlarda hareket eden büyük cisimlere göre çok daha zor olmaya devam ediyor. Bu tip nesnelerin takibi için Kalman filtresi, parçacık filtresi, bulanık mantık ve Gaussian modellemesi gibi farklı metodlar uygulanmıştır. Fakat son on yıl içinde evrişimsel sinir ağları kullanan yeni metodlar bu klasik metodlara alternatif olarak ortaya çıkmış ve birçok alanda büyük bir başarıyla uygulanmışlardır. Bu çalışmada, evrişimsel sinir ağlarını kullanarak pinpon topu gibi doğrusal olmayan yönlerde ve yüksek hızlarda hareket eden küçük cisimlerin gerçek zamanlı tespiti ve yardımcı algoritmalarla nesne takibinin yapılabilmesini sağlayan bir sistem geliştirilmiştir. Evrişimsel sinir ağı temelli bir nesne tespit algoritması olan YOLO, pinpon topunun farklı renk ve video üzerinde değişen biçimleriyle birlikte veri seti hazırlanarak eğitilmiş ve test edilmiştir. Büyük oranda başarı sağlandığı görülmüş, bu tür nesnelerin evrişimsel sinir ağı temelli algoritmalar ile gerçek zamanlı tespitinin ve takibinin mümkün olduğu görülmüştür. İleride yapılacak olan çalışmalarda masa tenisi oynayabilecek bir robot için araştırma yapılması planmaktadır. Bu nedenle, tespiti ve yardımcı algoritmalarla takibi yapılan pinpon topunu bir işaretçi ile sürekli olarak işaret eden 2 eksenli servo motor kullanan bir robot kol inşaa edilmiştir.
Computer vision technology has begun to develop since the 1960s and has made significant advances in this field today. Today, on real-time images, cars can be traced by objects that are in different movements like humans. But the detection and tracking of small objects that move very quickly in non-linear orbits continues to be much more difficult for most cases than big objects that move at more linear, normal speeds. For the tracking of this type of objects, different methods have been applied, such as Kalman filtration, particle filtration, foolish logic and Gaussian modeling. But in the last decade, new methods using evolutionary nerve networks have emerged as an alternative to these classic methods and have been applied with great success in many areas. In this study, a system was developed that uses evolutionary nerve networks to allow the detection in real-time of small objects that move in non-linear directions and at high speeds, such as the pinpon ball, and to help the objects to be traced with assistant algorithms. YOLO, an evolutionary nerve network-based object detection algorithm, has been trained and tested by preparing a set of data along with the varied forms of the pinpon ball on different colors and videos. It has been shown to have been achieved a large degree of success, and it has been shown that such objects are possible in real-time detection and tracking with algorithms based on the evolutionary nerve network. In the future, research is planned for a robot that can play table tennis. Therefore, a robot arm is built using a 2-axis servo engine that continually marks the pinpon ball with a mark, which is detected and traced by assisting algorithms.
Computer vision technology has been constantly evolving from 1960’s on, and a lot of progress has been made in this field since that date. Today, different moving objects such as cars and people can be tracked on real-time images. However, the detection and tracking of small objects moving in nonlinear trajectories and at high speeds is still more challenging compared to large objects moving in more linear and slower speeds. Previously, different methods such as Kalman filters, particle filters, fuzzy logic and Gaussian modeling have been applied to the problem tracking such objects. But, in the last ten years, new methods using convolutional neural nets emerged as alternatives to these classical methods and they were applied to various problems with great success. In this study, a convolutional neural network-based system will be developed which enables the real-time detection and tracking of a fast-moving small object, such as a ping-pong ball. YOLO, which is a neural network-based object detection algorithm, is trained on the images of fast-moving ping-pong balls of various colors, where all the distortions accompanying fast moving objects like motion blur are present. For this purpose, a new training set is created. A high success rate is achieved in detecting and tracking the ping-pong ball, and it has been observed that real-time detection of such objects is possible with convolutional neural network-based algorithms. In future, this research is planned to be extended to a robot that can play table tennis. For this purpose, a two degree of freedom robot arm using 2 servos has been built, which continuously points the ping-pong ball with a pointer as it travels.
Field : Fen Bilimleri ve Matematik; Mühendislik
Journal Type : Uluslararası
Article Detail 
Citation Owners
Similar Articles
Download PDF
Journal Information
Listen Paper
Print This Page
Share
Mendeley
