覃业宝,孙 炜,范诗萌,张 星,刘 剑.全距离深度平衡立体匹配网络[J].电子测量与仪器学报,2023,37(8):30-39 |
全距离深度平衡立体匹配网络 |
Full range depth balanced stereo matching network |
|
DOI: |
中文关键词: 立体匹配 深度代价体 视差与深度损失融合 7 邻域特征 视差优化 深度精度 |
英文关键词:stereo matching depth cost volume disparity and depth loss fusion seven neighborhood features disparity optimization depth accuracy |
基金项目:国家自然科学基金(U22A2059)、深圳科技计划项目(2021Szvup035)、湖南大学汽车车身先进设计制造国家重点实验室自主研究项目、电子制造业智能机器人技术湖南省重点实验室开放课题项目资助 |
|
Author | Institution |
Qin Yebao | 1. School of Electrical and Information Engineering,2. State Key Laboratory of Advanced Vehicle Design and Manufacturing, Hunan University |
Sun Wei | 1. School of Electrical and Information Engineering, Hunan University,3. Shenzhen Research Institute, Hunan University |
Fan Shimeng | 1. School of Electrical and Information Engineering,2. State Key Laboratory of Advanced Vehicle Design and Manufacturing, Hunan University |
Zhang Xing | 1. School of Electrical and Information Engineering, Hunan University,3. Shenzhen Research Institute, Hunan University |
Liu Jian | 1. School of Electrical and Information Engineering,2. State Key Laboratory of Advanced Vehicle Design and Manufacturing, Hunan University |
|
摘要点击次数: 656 |
全文下载次数: 599 |
中文摘要: |
针对当前视差估计网络在将视差转换成深度时,存在深度精度受相机参数影响,且在远距离处产生深度精度急剧下降
的问题,提出一种全距离深度平衡立体匹配网络(FRDBNet)。 首先构建深度代价体,使网络学习到全距离深度的概率分布,进
行深度回归直接生成深度;然后采用视差与深度损失融合的训练策略使网络同时关注远中近三分段全距离的深度估计;最后,
基于初始视差右图对应点 7 邻域特征设计视差优化模块进一步提高网络的深度估计精度。 在大型真实驾驶场景 DrivingStereo
数据集上的实验表明,针对全距离[1,100]m 的深度估计,FRDBNet 在[1,30]m 近距离、[30,60]m 中距离和[60,100]m 远距离
处深度精度相比 CVPR2022 性能表现优越的 ACVNet 分别提高 10. 38%、15. 11%和 20. 35%,达到了良好的深度精度平衡。 |
英文摘要: |
In view of the problem that depth accuracy is affected by camera parameters when disparity is converted into depth in the
current disparity estimation network, and depth accuracy decreases sharply at long distance, a full range depth balanced stereo matching
network (FRDBNet) is proposed. Firstly, the depth cost volume is constructed to make the network learn the probability distribution of
the full distance depth, and the depth is directly generated by depth regression. Then, the training strategy of disparity and depth loss
fusion is used to make the network pay attention to the depth estimation of the long, middle and near three segments distance at the same
time. Finally, a disparity optimization module is designed based on the seven neighborhood features corresponding to the original
disparity right map to further improve the depth estimation accuracy of the network. Experiments on the DrivingStereo dataset of large
real-world driving scenarios show that for the full distance[1,100]m depth estimation, the depth accuracy of FRDBNet at[1,30]m short
distance,[30,60]m middle distance and[60,100]m long distance is 10. 38%, 15. 11% and 20. 35% higher than that of ACVNet with
superior performance of CVPR2022, respectively, achieving a good balance of depth accuracy. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|