Color-Guided Depth Recovery Based on ZED

Depth maps obtained by RGB-D cameras such as ZED are noisy, and there are invalid areas where depth values absent in the maps, which will affect the subsequent improved processing. This paper investigates the restoration and optimization of degraded depth image method based on RGBD data captured by ZED by means of selective filtering and conditional auto-regressive model with edge constraints. Firstly, we divide invalid areas into two categories, including the internal invalid areas caused by reflection and invalid edge caused by occlusion. Then, we use corresponding median filter and mean filter in the different selective areas to obtain the initial restored depth map with no empty area. The edges of objects contain massive information, therefore, an adaptive auto-regressive (AR) model is used to further optimize the depth of the edge area. Additionally, we add the edge information to the sampled image to maintain local detail. Experiments show that with the proposed method we can recover the depth data with high accuracy and effectiveness.