The Dense Depth Estimation Algorithm For Wide-Angle Images

Wide-angle cameras can be used to perceive and reconstruct the surrounding environment in virtual reality and robotics.The state-of-art depth estimation algorithms can not generate accurate dense depth map for wide-angle images.In this dissertation,we aim to generate dense depth maps from wide-angle images.First,we propose to extend the two-view geometry of perspective projection images with the wide-angle lens projection model,so that the epipolar geometry and homography can be applied to wide-angle images.Second,we present a real-time camera pose tracking algorithm for wide-angle images under unified camera model.Third,we use Patch Match Stereo for wide-angle images with the adaptation of the camera projection model.The proposed algorithm can obtain high-quality depth maps.Besides,we implement a verification system for the dense depth estimation algorithm of wide-angle images,which contains the proposed algorithms and several other algorithms for comparison.This system can be used for algorithm analysis of wide-angle images.Our main contributions include·The geometry of two wide-angle images:This dissertation extends the epipolar geometry and planar homography to two wide-angle views with the projection models of the wide-angle lens.By adopting the wide-angle lens projection model to the two-view geometry,we can get the camera extrinsic directly from the wide-angle images without rectification.·Camera trajectory tracking for wide-angle image:In this dissertation,we propose to extend the wide-used ORB-SLAM system with the unified camera model.The proposed method first applies two views geometry of the wide-angle images to get the initial camera pose estimation.Then we apply the unified camera model to the bundle adjustment process for the optimization of reprojection error.By optimizing the initial pose,we can get an accurate wide-angle camera trajectory.·Dense depth estimation for wide-angle images:In this dissertation,we propose applying the wide-angle camera projection model to the Patch Match Stereo.Our main contribution includes following aspects.First,we apply the wide-angle camera projection model which enables the direct stereo matching on wide-angle images without rectification.Second,we use the homography to perform planar mapping for each pixel,so the proposed algorithm can deal with stereo pairs with an arbitrary baseline.Third,using patch-based stereo matching,our algorithm can locate correct planes for each pixel with a random search strategy,which avoids missing the accurate depth values.Our algorithm can directly estimate accurate depth values from the given wide-angle images.·Algorithms verification system for depth estimation for wide-angle images:Our system implements camera trajectory tracking and dense depth map estimation for wide-angle images.We also generate point clouds from continuous depth maps.Our system can be used for depth estimation algorithm analysis.We perform quantitative and qualitative comparisons against the state-of-the-art wide-angle image stereo matching methods.Experimental results show our algorithm can generate high-quality depth maps.