Abstract
Disparity estimation of binocular images is an important branch of depth reconstruction. The neural network method based on depth learning shows excellent performance in binocular stereo vision, which is far beyond the traditional binocular depth estimation algorithm. Most binocular stereo matching networks rely on training the dataset of the target scene, and can not predict the disparity of bottom components of the train directly. For this problem, we propose a domain-adaptive network, which includes two types of feature extraction encoding layers, a cost volume module that fuses a pair of image features, and a disparity prediction module based on multi-scale convolution layer. Moreover, based on the results of network prediction, we use a consistency checking method to further refine the disparity and enhance the reliability of the estimation results. Through experiments in different scenes, our method shows more accuracy and generalization than ordinary stereo matching networks, and has good effect of stereo visualization with less memory consumption in real train scenes. At the same time, our optimization method of consistency reduces the error of prediction and improves the confidence of disparity estimation for train bottom as a whole.
Funding Information
  • National Natural Science Foundation of China