Örtmeleri gözeterek O(1) karmaşıklıkta stereo eşleme.


Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2014

Tezin Dili: İngilizce

Öğrenci: Yeti Ziya Gürbüz

Danışman: ABDULLAH AYDIN ALATAN

Özet:

The problem of joint reduction of computational complexities of local stereo matching methods due to both cost aggregation step and correspondence search range is addressed and a novel hierarchical stereo matching algorithm is presented. The proposed approach exploits edge aware recursive volume filtering with a reduction on correspondence search range. The fastest state-of-the-art edge aware recursive filters are modified so that they become applicable to the methods to reduce the complexity in correspondence search range. In this way, complexities due to both cost aggregation step and correspondence search range are eliminated, yielding an O(1) complexity algorithm. In addition, the weakness of recursive filters in the presence of noise or high texture is handled by the help of a hierarchical scheme. Unlike common hierarchical methods, the transfer of the disparity estimates across the scales is converted into an optimization problem in order to preserve object boundaries and propagate proper estimates across scales. Dynamic programming is exploited to solve this optimization problem efficiently. The proposed transfer method can be utilized to transfer disparity across scales either in an image pyramid between stereo pairs or along frames in stereo video. The occlusion problem is solved inherently by the proposed approach that provides further decrease in complexity. The experimental results show that the proposed method provides quite efficient computation for stereo matching with a marginal decrease in performance. Compared to the state-of-the-art techniques, the proposed technique is possibly the fastest approach with a comparable accuracy based on benchmarking with Middlebury stereo pairs.