Skip to content
Open access

Lightweight Monocular Depth Estimation Based on Structural Fusion and Edge-Aware Distillation

Nov 2026 · IEEE Sensors Journal · Vol 26, pp. 33198-33211 · 0 citations · 48 references

Abstract

Self-supervised monocular depth estimation (MDE) eliminates the reliance on expensive ground-truth depth annotations and has emerged as a crucial technique for 3-D scene understanding in autonomous driving and robotic navigation. In this article, a lightweight self-supervised MDE model is proposed for balancing global context modeling and computational efficiency through structural fusion and edge-aware distillation. The model integrates an efficient model (EMO) as its backbone, effectively combining the local feature extraction capability of CNNs with the global reasoning ability of Transformers. To enhance feature representation in the decoder, a spatial-channel fusion attention (SC-FA) module is proposed, which improves cross-level feature fusion while preserving spatial relationships essential for depth coherence. Furthermore, an edge-aware knowledge distillation framework is established to transfer geometric priors from a pretrained MonoViT teacher model, enabling model compression without sacrificing accuracy. Experiments on the KITTI benchmark demonstrate the competitive performance of our method: our full model achieves an absolute relative error (Abs Rel) of 0.104 and a threshold accuracy of 0.893 with only 5.8 M parameters and $\delta \lt {1}.{25}$ , outperforming Monodepth2-Res18 (14.3 M, 0.115 Abs Rel, 0.883 accuracy with $\delta \lt {1}.{25}$ ). Furthermore, its scalable lightweight variants enable flexible complexity reduction, making it adaptable for real-time deployment on resource-constrained edge devices.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.