Pozlanmış görüntüleri sınıflandırarak etiketleme


Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Eskişehir Osmangazi Üniversitesi, FEN BİLİMLERİ ENSTİTÜSÜ, ELEKTRİK-ELEKTRONİK MÜHENDİSLİĞİ, Türkiye

Tezin Onay Tarihi: 2010

Tezin Dili: Türkçe

Öğrenci: TUĞRUL GÜÇLÜ

Danışman: Nihat Adar

Özet:

In this work, the process of recognition human actions such as handwaving,walking, runnig by using angular features of body parts is presented. In the first parts ofthe study a human body part model of 10 parts has been designed by using pictorialstructures. This model is again used in the manuel labeling process of images belongingto two different image database. At the end of this process, a library of angular featureshas been formed. Manuel labeling also creates an abstraction layer for image processingmethods which can automatically find body parts. Also working with different imagedatabases makes it possible to test usage of angular features against images with differentscales and orientations. After having angular features, k-means clustering is used togather pose words from angular features of body parts. In order to recognize actionsfrom pose words SVM and Neural Network classifers are used. Results are thencompared whether common pose or repetitive pose appending method is used to equalizeinput vector lenghts. In addition to these tests, other tests are applied to understand theimportance of pose arrangement order in action recognition. These test involve shufflingand reversing the pose arrangement order. Results also imply the importance of posearrangement order in action recognition.