DispNet based Stereo Matching for Planetary Scene Depth Estimation Using Remote Sensing Images

Recent work has shown that convolutional neural network can solve the stereo matching problems in artificial scene successfully, such as buildings, roads and so on. However, whether it is suitable for remote sensing stereo image matching in featureless area, for example lunar surface, is uncertain. This paper exploits the ability of DispNet, an end-to-end disparity estimation algorithm based on convolutional neural network, for image matching in featureless lunar surface areas. Experiments using image pairs from NASA Polar Stereo Dataset demonstrate that DispNet has superior performance in the aspects of matching accuracy, the continuity of disparity and speed compared to three traditional stereo matching methods, SGM, BM and SAD. Thus it has the potential for the application in future planetary exploration tasks such as visual odometry for rover navigation and image matching for precise landing