基於多視角影像之戶外場景前景背景分離技術研究

Abstract

在影像分割的技術中加入少量的使用者自定義的前後景仍然難以在複雜環境下切割出完美的前景物，而對於全自動的前景物切割，建立背景模型便是我們的第一步。在本篇論文中，我們著重在即使在戶外的複雜環境，仍然可以用低成本的系統去自動的達到前景背景分離的技術。首先，基於混合高斯背景模型，再配合非線性色彩分佈和多視角影像的輔助，我們設計最佳化問題來得到初步的前景物資訊，並且抵抗影子的影響，之後再利用影像分割的概念配合背景的資訊和初步的前景資訊去得到最後切割完整的前景物。有了切割出來的前景物，再配合人為的合成背景便可以在戶外輕鬆得到室內虛擬攝影棚想要的效果。實驗結果證明我們所提出的方法有許多前景分割的好處，其中包含全自動化、移除影子的干擾和即便在戶外環境中也能得到完整而且邊緣平滑的前景物。

The state-of-art interactive image segmentation algorithms often have difficulty in correctly extracting the foreground objects from cluttered background with limited user’s guidance. For automatic foreground object detection, constrained background models are critical. In this thesis, we propose a low-cost automatic foreground/background separation system that can be applied to outdoor scenes. Based on a Gaussian mixture model, together with the inclusion of non-linear tone mapping and multi-view image constraint to further eliminate shadow effect, we formulate an optimization problem to deal with the foreground extraction problem by using more robust image features in the image matting technique. The proposed method exhibits many desired properties of an effective foreground segmentation algorithm, including automatically extraction of foreground regions, the ability to produce smooth and accurate boundary contour, and the ability to handle severe color variations in an outdoor environment with relaxed background constraints. The whole system can achieve fully automatic foreground object extraction with satisfactory accuracy for a multi-camera system.