Yüksek başarımlı, bigisayarla görü uygulamaları programlaması


Tezin Türü: Doktora

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

Tezin Onay Tarihi: 2009

Tezin Dili: Türkçe

Öğrenci: CELAL MURAT KANDEMİR

Danışman: Nihat Adar

Özet:

Human action detection in video sequence and recognition of these actions aremajor working fields in computer vision. Action detection and recognition are used inmany areas like education, security, and military. In this thesis, human activity detectionand recognition systems based on Artificial Neural Networks (ANN) and HiddenMarkov Model (HMM) are modeled for detection and recognition of the humanactivities from video sequences. The action recognition system models consist of threestages; human pose detection, pose sequence generation, and action recognition fromthese pose sequences. At the first video frame, an iterative algorithm using edge baseddeformable model is employed to detect human pose. Afterwards, this algorithmcombined with the three step search method is used to find human pose in the remainingframes. With this method, pose finding times are improved. At the second stage, allfeature vectors obtained from the video sequences are fed into classification algorithm.Consequently, pose symbol repetition is prevented and new pose symbols that are not inthe codebook can be easily added. The pose sequences are obtained by the vectorquantization. Addition of repeated and common pose operations are defined as twodifferent normalization methods for the action sequences of different lengths. At thethird stage, codebooks of different lengths and normalized pose sequences are used fortraining and testing ANN, and HMM based recognition models. Common pose addition reduces the learning time and provides high recognition rates in ANN models.