EUVIMEDEUVIMED entdecken
Crossref

Self-Supervised Monocular Depth Estimation With Multimodal Fusion and Pose Enhancement

Han Chen, Yongxiong Wang, Shuwen Jia, Shuai Huang, Jiayi Zhang

IEEE Sensors Journal · 01.06.2026 · 26 (11) · 17339-17355

DOI: 10.1109/jsen.2026.3684347 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 01.06.2026. Verfasst von Han Chen und 4 weiteren Autor:innen. Die Arbeit nennt 62 Literaturangaben. Gefördert von: Natural Science Foundation of Shanghai.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The application of self-supervised learning to monocular depth estimation eliminates the need for labor-intensive depth annotations, thus unlocking the potential of large-scale unlabeled video data and establishing itself as a prominent research focus in computer vision. However, its performance is often limited by three critical challenges: 1) scale ambiguity caused by camera pose uncertainty, which fundamentally impacts depth estimation accuracy; 2) performance limitations in depth estimation networks; and 3) perturbations caused by dynamic objects and textureless regions. To address these challenges, we propose a multimodal fusion pose-enhanced network (MFPE-Net), a multimodal fusion pose-enhanced self-supervised monocular depth estimation network. First, we introduce a visual–inertial enhanced pose generator (VIEPG) that integrates absolute scale information from inertial data with relative pose estimates from images, significantly improving pose estimation accuracy and robustness. Second, we design the hierarchical query-driven energy-aware depth estimator (HQEDE), which employs a progressive query refinement mechanism to efficiently capture the intrinsic relationships between pixels and instances for fine-grained depth estimation. Finally, we develop an innovative masking strategy, the reverse feature-metric bundle adjustment (RFBA) mask, to effectively identify and mitigate the impact of dynamic objects and textureless regions on depth estimation. Extensive experiments on the KITTI and Cityscapes datasets demonstrate that our method consistently outperforms state-of-the-art approaches across multiple evaluation metrics.

Förderung

  • Natural Science Foundation of Shanghai

Zitieren

APA

Chen, H., Wang, Y., Jia, S., Huang, S., & Zhang, J. (2026). Self-Supervised Monocular Depth Estimation With Multimodal Fusion and Pose Enhancement. IEEE Sensors Journal, 26(11), 17339-17355. https://doi.org/10.1109/jsen.2026.3684347

Vancouver

Chen H, Wang Y, Jia S, Huang S, Zhang J. Self-Supervised Monocular Depth Estimation With Multimodal Fusion and Pose Enhancement. IEEE Sensors Journal. 2026;26(11):17339-17355. doi:10.1109/jsen.2026.3684347

BibTeX
@article{Chen2026,
  author = {Chen, Han and Wang, Yongxiong and Jia, Shuwen and Huang, Shuai and Zhang, Jiayi},
  title = {Self-Supervised Monocular Depth Estimation With Multimodal Fusion and Pose Enhancement},
  journal = {IEEE Sensors Journal},
  year = {2026},
  volume = {26},
  number = {11},
  pages = {17339--17355},
  doi = {10.1109/jsen.2026.3684347}
}
Kostenlos starten →