Volltext-Downloads (blau) und Frontdoor-Views (grau)

Improving the Depth Estimation of Stereo Cameras through Deep Learning-Based Image Denoising

  • Stereo depth cameras such as the RealSense D455 are low-cost and widely deployed, but their accuracy is limited by image noise, with depth error growing quadratically with distance. Replacing consumer-grade sensors with higher-precision hardware is often infeasible due to cost, size, and power constraints. This thesis therefore investigates a software-only, deep-learning-based denoiser that improves depth accuracy without hardware change. A compact U-Net operating on inverse depth, trained with residual learning and a masked BerHu loss, is selected over Transformer- and State-Space-Model-based alternatives for its accuracy, efficiency, and simplicity. Training and evaluation draw on four self-recorded RealSense datasets. Among them is an independent Test-LiDAR set with ground truth from an Ouster OS1 LiDAR, which probes generalization beyond the temporally averaged pseudo-ground truth used elsewhere. A grid search yields two Pareto-optimal variants, U-Net-Base (1.93M parameters) and U-Net-Lite (466k parameters), which outperform three classical denoisers (bilateral filter, guided filter, BM3D) on every disparity metric. U-Net-Base reduces disparity RMSE by 27.6% and improves PSNR by 2.89 dB on the in-domain test set, versus 6.8%/0.76 dB on the harder, out-of-domain Test-LiDAR set. Both variants reach 350–400 FPS on a workstation GPU and 35–40 FPS on a consumer-grade GTX 1080, meeting real-time requirements. The denoiser cannot recover missing pixels or correct large detection errors. Its smaller Test-LiDAR gains suggest that dataset diversity, not model capacity, is the main bottleneck to further improvement. These results show that a lightweight, software-only denoiser can narrow the accuracy gap to costlier depth-sensing hardware.

Download full text files

Export metadata

Statistics

frontdoor_oas
Metadaten
Author:Jannis Schaake
URN:urn:nbn:de:bsz:960-opus4-40167
DOI:https://doi.org/10.25968/opus-4016
Advisor:Volker AhlersORCiDGND, Adrian PigorsGND
Document Type:Master's Thesis
Language:English
Year of Completion:2026
Publishing Institution:Hochschule Hannover
Granting Institution:Hochschule Hannover, Fakultät IV - Wirtschaft und Informatik
Date of final exam:2026/09/18
Release Date:2026/10/12
Tag:Convolutional Neural Network; Depth Estimation; Depth Map; Image Denoising; Stereo Vision
GND Keyword:StereokameraGND; TiefenerkennungGND; BildrauschenGND; Deep learningGND; LidarGND
Page Number:125
Institutes:Fakultät IV - Wirtschaft und Informatik
DDC classes:004 Informatik
Licence (German):License LogoCreative Commons - CC BY - Namensnennung 4.0 International