新型遮擋填充與Canny邊界分割檢測相結合的立體深度恢復方法

Abstract

立體視覺是電腦視覺中的一個重要組成部分，也一直是幾十年來研究的熱點。雙目立體視覺實際上是模仿人類視覺獲取深度資訊以及三維場景重建的過程，其應用範圍從機器人導航、工業測量到醫療和軍事等方面，獲取密集的深度圖是本文的研究重心。 到目前為止，遮擋區域，深度不連續區域，弱紋理等一系列問題是獲取精確深度圖的主要障礙。在本文中我們提出了一種新的方法，它結合了色彩，空間和圖像分割等資訊來填充遮擋的無效圖元，並用同樣方式作用於整個圖像像素從而保證深度圖一致性和完整性。我們工作的另一個創新是邊緣恢復機制的引入，用來處理深度不一致的區域，隨後的雙邊濾波和平滑處理進一步提高最終的深度圖的品質。我們在Middlebury dataset平臺上測試了我們的演算法並取得了顯著的效果。我們還將演算法對真實世界的室內和室外的圖像進行測試，結果表明我們的演算法在不同的條件下取得了不錯的深度圖也驗證了演算法的魯棒性。

Stereo vision is an important part of computer vision and has been research hotspot for decades. Binocular stereo vision is actually the process of mimicking human vision to obtain depth information and reconstruct 3D scenes, its application ranges from robot navigation, industrial measurement to medical treatment and military affairs, acquiring dense accurate depth maps is the main concern in our paper. So far, the technical problems of occlusion regions, depth inconsistent and weak texture are the main obstacles in gaining accurate depth maps. Our paper propose an novel method which combines the elements of color, spatial and image segmentation information to fill the occluded pixels and the same principle is applied to the whole image pixels for consistency and integrity. Another innovation of our work is the introduction of an edge restoration mechanism which performs well in dealing with depth inconsistent region, subsequent bilateral filter and smoothing processing further improve the quality of final depth maps. We test our algorithms on Middlebury dataset and have remarkable results. Test on real-world sources of indoor and outdoor images indicates that our algorithm has good robustness, capable of gaining decent dense depth maps under various conditions.