摘要
当前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,造成图像在不适定区域的特征提取能力较差,因此,提出了一种基于PSMNet改进的优化立体匹配算法。在特征提取阶段,全新的特征金字塔模块(SPP)能更好的聚合不同尺度和不同位置的环境信息构建代价体,从而充分利用全局环境信息;在构建匹配代价体时,提出组相关的策略来充分地利用特征中的全局和局部信息;在代价聚合阶段,优化沙漏结构并引入通道注意力机制以便网络来提取具有高表示能力和高质量通道注意力向量的信息特征;为了进一步优化视差图,设计视差优化网络来改善初始的视差估计。在Scene Flow、KITTI 2012和KITTI 2015立体数据集上评估,所提模型在Scene Flow数据集上平均预测误差EPE降低到0.71 pixels,在KITTI 2012和KITTI 2015立体数据集上的误匹配率分别下降到1.20%和1.86%,在实验结果表明,方法取得了较优越的性能。
This paper presents an improved stereo matching algorithm based on PSMNet.In the feature extraction stage,the new SPP feature pyramid module can better aggregate the environmental information of different scales and different locations to construct cost volume,in order to make full use of the global environmental information.When constructing the matching cost volume,the group correlation strategy is proposed to make full use of the global and local information in features.In the cost aggregation stage,the hourglass structure is optimized and the channel attention mechanism is introduced so that the network can extract the information features with high representation ability and high quality channel attention vector.In order to further optimize the disparity map,a disparity optimization network is designed to improve the initial disparity estimation.The method in this paper is evaluated on Scene Flow,KITTI 2012 and KITTI 2015 stereo datasets,and the average prediction error EPE of the proposed model on Scene Flow dataset is reduced to 0.71 pixels.The mismatching rates on KITTI 2012 and KITTI 2015 stereo datasets decreased to 1.20%and 1.86%,respectively.The experimental results show that the proposed method achieves superior performance.
作者
谢鑫
张博
张美灵
朱磊
Xie Xin;Zhang Bo;Zhang Meiling;Zhu Lei(School of Electronics and Information,Xi'an Polytechnic University,Xi'an 710048,China)
出处
《国外电子测量技术》
北大核心
2023年第1期89-99,共11页
Foreign Electronic Measurement Technology
基金
国家自然科学基金(61971339)
陕西省重点研发计划(2019GY-113)
陕西省自然科学基础研究计划(2019JQ-361)项目资助
关键词
立体匹配
深度学习
注意力机制
卷积神经网络
分组相关量
视差优化
stereo matching
deep learning
attention mechanism
convolution neural network
group correlation quantity
parallax optimization
