基于卷积神经网络的立体匹配算法在精度上取得了较大的提高,但大多数算法仍然无法满足实时性要求。提出一种渐进细化的实时立体匹配算法,在低分辨率层级中初始化视差图,再渐进地恢复视差图的空间分辨率。该算法采用轻量的骨干网络提取多尺度特征,在保证算法实时性的同时,对特征进行反向融合,提高了特征的稳健性。提出一种多分支融合模块对视差图进行渐进细化,对不同区域的多种模式进行自动聚类,再分别预测视差图残差,根据聚类权重融合最终结果,使模型能够更好地处理具有不同特点的区域。在KITTI测试集上,所提算法的运行速度达到20 frame/s,与运行效率相当的DispNetC算法相比,错误率降低了约30%。

The stereo matching algorithm based on convolutional neural network improves the accuracy considerably; however, most algorithms still cannot meet the real-time requirements. In this study, we propose a real-time stereo matching algorithm with hierarchical refinement, which initializes the disparity map at a low-resolution level and gradually restores the spatial resolution of the disparity map. The proposed algorithm uses a lightweight backbone network for extracting the multi-scale features and simultaneously, the features are inversely fused to achieve an improved robustness without significantly affecting its real-time performance. Furthermore, we propose a multi-branch fusion module to progressively refine the disparity map. After the different modes in different regions are automatically clustered, the residuals of the disparity map are predicted. Subsequently, the final results are combined based on the cluster weights to ensure that the regions exhibiting different characteristics can be effectively processed. Based on the KITTI test dataset, the operating rate obtained using the proposed algorithm is 20 frame/s, and the error rate is reduced by approximately 30% when compared with the DispNetC algorithm. These exhibit a comparable operating efficiency.