摘要： 现有的立体匹配算法通常采用深层卷积神经网络提取特征，对前景物体的检测更加精细，但对背景中的小物体及边缘区域匹配效果较差。为提高视差估计质量，构建一个基于视差优化的立体匹配网络CTFNet。分别提取浅层与深层特征，并基于深层特征构建全局稀疏代价卷，从而预测初始视差图。在预测的初始视差图和浅层特征的基础上构建局部稠密代价卷并进行视差优化，以细化预测视差值邻域的概率分布，提高特征不明显区域的匹配精度。此外，引入新的概率分布损失函数，监督softmax函数计算的视差值概率分布在真实视差值附近成单峰分布，提高算法的鲁棒性。实验结果表明，该网络在SceneFlow和KITTI数据集上的误匹配率分别为0.768%和1.485%，在KITTI测评网站上的误差率仅为2.20%，与PSMNet网络相比，精度和速度均得到一定提升。

关键词: 立体匹配, 视差优化, 浅层特征, 匹配代价卷, 损失函数

Abstract: Existing stereo matching algorithms usually use deep convolutional networks to extract features, and can improve the accuracy of foreground object detection, but display poor matching results for small objects and boundary areas in the background.In order to improve the quality of disparity estimation in these areas, a stereo matching network named Coarse To Fine Net(CTFNet) is proposed based on disparity optimization.The network extracts shallow and deep features separately and a global sparse cost volume is constructed based on the deep features to predict the initial disparity map.Then a local dense cost volume is constructed based on the predicted initial disparity map and shallow features, which optimizes the disparity and refines the probability distribution of the neighborhood of predicted disparity value to improve the matching accuracy of areas with less obvious features.At the same time, a new loss function for probability distribution is introduced to supervise the probability distribution calculated by the softmax function in a unimodal distribution near the true disparity value and improve the robustness of the algorithm.The experimental results show that the mismatching rate of the proposed network is 0.768% on the SceneFlow dataset and 1.485% on the KITTI data set, and its error rate on the KITTI evaluation website is only 2.20%.Compared with the PSMNet network, the proposed algorithm displays an improvement in both accuracy and speed.

Key words: stereo matching, disparity optimization, shallow feature, matching cost volume, loss function

中图分类号: 

• TP391