一种基于PSMNet改进的立体匹配算法

Improved Stereo Matching Algorithm Based on PSMNet

为了解决双目视觉中的立体匹配问题、减少立体匹配网络的参数数量、降低算法的计算复杂度、提高算法的实用性。在PSMNet立体匹配网络的基础上进行改进,提出了一种具备浅层结构与宽阔视野的立体匹配算法--SWNet。浅层结构表示网络层数更少、参数更少、处理速度更快;宽阔视野则表示网络的感受野更宽广,能够获取并保留更多的空间信息。SWNet由特征提取、3D卷积和视差回归3个部分构成。在特征提取部分,引入了深色空间金字塔结构(Atrous Spatial Pyramid Pool,ASPP),用于提取多尺度的空间特征信息,设计了特征融合模块,将不同尺度的特征信息有效地融合起来以构建匹配代价卷;3D卷积神经网络利用堆叠的编码解码结构进一步对匹配代价卷进行规则化处理,获得不同视差条件下特征点之间的对应关系;最后,采用回归的方式得到视差图。SWNet在SceneFlow和KITTI 2015两个公开的数据集上均取得了优异的表现,与参考算法PSMNet相比,参数数量下降了48.9%,且误匹配率仅有2.24%。

Based on PSMNet stereo matching network,an improved stereo matching algorithm with shallow structure and wide receptive field-SWNet was proposed,in order to solve the stereo matching problem in binocular vision,reduce the number of parameters of the stereo matching network,reduce the computational complexity of the algorithm,and improve the practicability of the algorithm.The shallow structure means fewer layers,fewer parameters and faster processing speed,while wide receptive field means that the network is more receptive and can acquire and retain more spatial information.SWNet consists of three parts:feature extraction,3 D convolution and disparity regression.In the aspect of feature extraction,Atrous Spatial Pyramid Pool(ASPP)was introduced,which was used to extract multi-scale feature information.Feature fusion module was designed to fuse multi-scale feature information and build matching cost volume.The 3 D convolutional neural network use the stack encoding and decoding structure to further regularize the matching cost volume and obtain the corresponding relationship between the feature points under different disparity conditions.Finally,the disparity map was obtained by regression.SWNet performed well on both SceneFlow and KITTI 2015 public datasets,with a 48.9%reduction in the number of parameters and a 2.24%mismatching rate compared to the reference algorithm PSMNet.