基于多尺度和多层级特征融合的立体匹配方法

Stereo Matching Method Based on Multi-Scale and Multi-Level Features Fusion

基于卷积神经网络的立体匹配方法未充分利用图像中各个层级的特征图信息,造成对图像在不适定区域的特征提取能力较差。提出一种融合多尺度与多层级特征的立体匹配方法。通过在双塔结构卷积神经网络模型的前端设计一个池化金字塔层,提取图像的多尺度低层结构特征。在该网络模型的后端融合最后三层网络的高级语义特征来提取图像特征,并对图像特征进行相似性度量后输出视差图。在KITTI 2015数据集上的实验结果表明,与LUO和Anita方法相比,该方法的像素误差精度分别由14.65%、8.30%降至8.02%,且可得到细节信息更好的视差图。

The stereo matching method based on Convolutional Neural Network(CNN)does not make full use of the feature map information of each level in the image,resulting in the poor feature extraction performance in the ill posed region in the image.This paper proposes a stereo matching method based on multi-scale and multi-level features.A pooled pyramid layer is designed at the front end of the CNN model with double tower structure to extract the multi-scale low-level structural features of the image.In the back end of the network model,the high-level semantic features of the last three layers of network are fused to extract image features,and the disparity map is output after similarity measurement of image features.The experimental results on KITTI 2015 dataset show that compared with the LUO and Anita methods,the proposed method reduces the pixel error accuracy from 14.65%and 8.30%to 8.02%,and can obtain a disparity map with better detail information.