融合多核分组卷积与自适应特征的轻量化单目深度估计模型

doi:10.13474/j.cnki.11-2246.2026.0810

测绘通报 ›› 2026, Vol. 0 ›› Issue (8): 67-73,81.doi: 10.13474/j.cnki.11-2246.2026.0810

• 学术研究 • 上一篇

融合多核分组卷积与自适应特征的轻量化单目深度估计模型

刘瑶¹, 黄鹤¹, 马延杰¹, 马朝威², 任博洋¹, 杨军星¹

1. 北京建筑大学测绘与城市信息空间学院, 北京 102616;
2. 航天规划设计集团有限公司, 北京 100070

收稿日期:2025-12-08 发布日期:2026-09-12
通讯作者: 杨军星。E-mail：yangjunxing@bucea.edu.cn
作者简介:刘瑶(2001—)，女，硕士生，主要从事深度估计、计算机视觉等相关研究。E-mail：2108570424065@stu.bucea.edu.cn
基金资助:
国家自然科学基金(42201483)；北京建筑大学青年教师科研能力提升计划(X23001)

Lightweight monocular depth estimation model based on multi-kernel grouped convolution and adaptive feature fusion

Liu Yao¹, Huang He¹, Ma Yanjie¹, Ma Chaowei², Ren Boyang¹, Yang Junxing¹

1. School of Geomatics and Urban Spatial Information, Beijing University of Civil Engineering and Architecture, Beijing 102616, China;
2. China Aerospace Planning and Design Group Co., Ltd., Beijing 100070, China

Received:2025-12-08 Published:2026-09-12

摘要/Abstract

摘要： [目的] 针对自监督单目深度估计远距离物体边缘细节提取不足、深度整体一致性表达较弱及难以满足自动驾驶对实时三维感知需求等问题，本文提出了一种轻量化自监督单目深度估计模型。[方法] 首先设计了多核分组深度可分离卷积模块，实现局部细节与全局结构信息的有效融合，大幅增强特征表达能力；然后提出了一种注意力引导的自适应特征融合模块，以动态调整多尺度特征的权重分配，增强深度预测的整体一致性。[结果] 试验结果表明，与经典算法Monodepth2相比，本文方法参数量减少10.4%、计算量减少34.7%，在关键性能指标Abs Rel和δ₁上分别达到0.112和0.878，并在跨场景测试中展现出更优的泛化性能。[结论] 本文方法在轻量化设计下能有效提升深度估计精度，为自动驾驶实时三维感知提供了可靠方案。

关键词: 单目深度估计, 自动驾驶, 自监督, 自适应特征融合, 多核分组卷积, 深度可分离

Abstract: [Purposes] To address the problems of insufficient extraction of distant object edge details，weak global consistency of depth prediction，and the difficulty of meeting real-time 3D perception requirements in autonomous driving for self-supervised monocular depth estimation，this paper proposes a lightweight self-supervised monocular depth estimation model. [Methods] Firstly，a multi-kernel grouped depthwise separable convolution module is designed to effectively fuse local detail and global structural information，significantly enhancing feature representation capability.In addition，an attention-guided adaptive feature fusion module is introduced to dynamically adjust multi-scale feature weighting and enhance the overall consistency of depth prediction. [Findings] Experimental results show that，compared with classical methods Monodepth2，the proposed approach reduces parameters by 10.4% and computation by 34.7%，achieving 0.112 and 0.878 on key performance metrics Abs Rel and δ₁，respectively，and demonstrating superior generalization performance in cross-scene testing. [Conclusions] The proposed method effectively improves depth estimation accuracy under a lightweight design，providing a reliable solution for real-time 3D perception in autonomous driving．

Key words: monocular depth estimation, autonomous driving, self-supervised learning, adaptive feature fusion, multi-kernel grouped convolution, depth separable

中图分类号:

P237

刘瑶, 黄鹤, 马延杰, 马朝威, 任博洋, 杨军星. 融合多核分组卷积与自适应特征的轻量化单目深度估计模型[J]. 测绘通报, 2026, 0(8): 67-73,81.

Liu Yao, Huang He, Ma Yanjie, Ma Chaowei, Ren Boyang, Yang Junxing. Lightweight monocular depth estimation model based on multi-kernel grouped convolution and adaptive feature fusion[J]. Bulletin of Surveying and Mapping, 2026, 0(8): 67-73,81.

导出引用管理器 EndNote|Reference Manager|ProCite|BibTeX|RefWorks

链接本文: http://tb.chinasmp.com/CN/10.13474/j.cnki.11-2246.2026.0810

http://tb.chinasmp.com/CN/Y2026/V0/I8/67

参考文献

[1] 胡海洋，陈超平，高天沐，等.单/双目深度估计研究进展与应用综述[J].红外与激光工程，2025，54(7):25-38．
[2] 郭迟，刘阳，罗亚荣，等.图像语义信息在视觉SLAM中的应用研究进展[J].测绘学报,2024，53（6）：1057-1076.
[3] Ranjan A，Jampani V，Balles L，et al.Competitive collaboration:joint unsupervised learning of depth，camera motion，optical flow and motion segmentation[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).Long Beach：IEEE，2020:12232-12241．
[4] Wang Chaoyang，Buenaposada J M，Zhu Rui，et al.Learning depth from monocular videos using direct methods[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Salt Lake City：IEEE，2018:2022-2030．
[5] 蒋祥龙，邓文亮，何胜喜.单目视觉驱动的机器人实时高精度稠密场景重建算法[J].测绘通报，2025(10):71-75．
[6] Casser V，Pirk S，Mahjourian R，et al.Depth prediction without the sensors:leveraging structure for unsupervised learning from monocular videos[J].Proceedings of the AAAI Conference on Artificial Intelligence，2019，33(1):8001-8008．
[7] Sun Libo，Bian Jiawang，Zhan Huangying，et al.SC-DepthV3:robust self-supervised monocular depth estimation for dynamic scenes[J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2024，46(1):497-508．
[8] Zhou Zhongkai，Fan Xinnan，Shi Pengfei，et al.R-MSFM:recurrent multi-scale feature modulation for monocular depth estimating[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision.Montreal：IEEE，2022:12757-12766．
[9] Zou Yuliang，Luo Zelun，Huang Jiabin.DF-Net:unsupervised joint learning of depth and flow using cross-task consistency[C]//Proceedings of 2018 Computer Vision-ECCV.Cham:Springer，2018:38-55．
[10] Luo Chenxu，Yang Zhenheng，Wang Peng，et al.Every pixel counts:joint learning of geometry and motion with 3D holistic understanding[J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2020，42(10):2624-2641．
[11] Godard C，Mac Aodha O，Firman M，et al.Digging into self-supervised monocular depth estimation[C]//Proceedings of 2019 IEEE/CVF International Conference on Computer Vision.Seoul：IEEE，2019:3827-3837．
[12] Han D，Shin J，Kim N，et al.TransDSSL:transformer based depth estimation via self-supervised learning[J].IEEE Robotics and Automation Letters，2022，7(4):10969-10976.
[13] Zhang Ning，Nex F，Vosselman G，et al.Lite-mono:a lightweight CNN and transformer architecture for self-supervised monocular depth estimation[C]//Proceedings of 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Vancouver：IEEE，2023:18537-18546．
[14] Zhang Jiacheng，Wang Xichao，Wang Xinyao，et al.CT-mono:leveraging CNNs and transformers for self-supervised depth estimation in single-view scenarios[J].Arabian Journal for Science and Engineering，2026，51(5):6023-6038．
[15] Bae J，Moon S，Im S.Deep digging into the generalization of self-supervised monocular depth estimation[J].Proceedings of the AAAI Conference on Artificial Intelligence，2023，37(1):187-196．
[16] Dai Yimian，Gieseke F，Oehmcke S，et al.Attentional feature fusion[C]//Proceedings of 2021 IEEE Winter Conference on Applications of Computer Vision.Waikoloa：IEEE，2021:3559-3568．
[17] Eigen D，Fergus R.Predicting depth，surface normals and semantic labels with a common multi-scale convolutional architecture[C]//Proceedings of 2015 IEEE International Conference on Computer Vision.Santiago：IEEE，2016:2650-2658．

融合多核分组卷积与自适应特征的轻量化单目深度估计模型

Lightweight monocular depth estimation model based on multi-kernel grouped convolution and adaptive feature fusion

PDF

可视化

摘要/Abstract

引用本文

使用本文

参考文献

相关文章 15

编辑推荐

Metrics

本文评价

[1]	王军强, 孙震辉, 孟庆岩, 张琳琳. 面向多云地区的置信度引导多模态Transformer水稻提取[J]. 测绘通报, 2026, 0(5): 44-49,79.
[2]	应申, 漆璇, 李玉, 王润泽, 鲁月新, 陶璐. 分段分层的立交桥建模方法[J]. 测绘通报, 2026, 0(1): 122-129.
[3]	张恒琦, 林茜楠, 李樵民, 闫亚亚. 基于隐马尔可夫模型的高精度地图路口模型自动化构建[J]. 测绘通报, 2025, 0(7): 97-103.
[4]	马小龙, 马照亭, 赵园春, 方驰宇, 费雯凯, 闫春利, 王月明, 石一慧. 自动驾驶地图在线更新安全合规技术研究与实践[J]. 测绘通报, 2025, 0(11): 134-139.
[5]	吕轩凡, 柳景斌, 毛井锋. 一种毫米波雷达里程计自主定位技术[J]. 测绘通报, 2024, 0(7): 1-5.
[6]	谢幸生, 张永挺, 丁宗宝, 江玉欢, 刘剑. 激光导航与决策融合的小样本无人机航拍图像分类[J]. 测绘通报, 2024, 0(7): 173-177.
[7]	费雯凯. 面向高精地图的局部相对位置动态生成方法及其应用[J]. 测绘通报, 2023, 0(6): 1-5.
[8]	赵明姝. 车载导航到自动驾驶电子地图发展综述[J]. 测绘通报, 2023, 0(6): 6-10,92.
[9]	王丽妍, 周勋, 胡伟, 刘立歆, 李飞雪. 面向自动驾驶的高精地图数据引擎模型[J]. 测绘通报, 2023, 0(6): 11-14,19.
[10]	鲁维佳, 刘泽帅, 潘玉恒, 李国燕, 李慧洁, 丛佳. 基于循环特征融合的弯道增强车道线检测算法[J]. 测绘通报, 2023, 0(12): 25-30.
[11]	宋伟东, 毕春杨, 赵奉书. 改进Mask R-CNN公路病害检测算法[J]. 测绘通报, 2023, 0(1): 58-64.
[12]	黄鹤, 孟维明. 基于视觉的大半径圆曲线车道线识别[J]. 测绘通报, 2022, 0(4): 134-137.
[13]	杨振凯, 华一新, 訾璐, 张政, 王继伟. 浅析高精度地图发展现状及关键技术[J]. 测绘通报, 2021, 0(6): 54-60.
[14]	周继苗, 李必军, 陈世增. 一种多层特征融合的道路场景实时分割方法[J]. 测绘通报, 2020, 0(1): 10-15.
[15]	陈世增, 李必军, 周继苗. 利用双灭点估计的车道线检测[J]. 测绘通报, 2020, 0(1): 16-20.