| Title: |
Visual Floorplan Localization Based on BEV Perception |
| Authors: |
CHEN Jiwei, CHEN Zebin, TAN Guang |
| Source: |
Jisuanji kexue, Vol 53, Iss 1, Pp 216-223 (2026) |
| Publisher Information: |
Editorial office of Computer Science |
| Publication Year: |
2026 |
| Collection: |
Directory of Open Access Journals: DOAJ Articles |
| Subject Terms: |
bev perception|floorplan localization|visual localization|geometric-semantic joint matching; Computer software; QA76.75-76.765; Technology (General); T1-995 |
| Description: |
Visual floorplan localization task achieves pose estimation by matching visual observation with scene floorplan representation.In practical applications,how to effectively integrate geometric and semantic correlations between observation and floorplan in matching is particularly important for improving localization accuracy.However,existing methods have two main li-mitations.Firstly,they fail to fully utilize the semantic information within the camera’s field of view.Secondly,they lack a joint matching mechanism for geometric and semantic clues.To address these issues,this study proposes a visual floorplan localization framework based on BEV perception,which includes three core components.Firstly,the BEV semantic mapping module constructs the BEV semantic representation of local scenes through multimodal image projection transformation,achieving structured representation of observation data.Secondly,the expected observation generation module generates an expected observation database within the floorplan space,and achieves rapid generation of observation data through differentiable rendering method.Finally,the multi-level matching and localizing module proposes a geometric-semantic joint matching mechanism,which integrates the geometric layout and semantic category information from BEV observations through a hierarchical matching strategy to achieve accurate matching with the floorplan.The experimental results show that the framework achieves an improvement in localization recall from 0.32% and 4.82% to 3.12% and 58.77% on the publicly available dataset Structured3D and the self built simulation environment dataset IndoorEnv,respectively,which is significantly better than the existing baseline methods Laser and F3Loc.This validates the effectiveness and robustness of the proposed method in complex indoor scenes. |
| Document Type: |
article in journal/newspaper |
| Language: |
Chinese |
| Relation: |
https://www.jsjkx.com/fileup/1002-137X/PDF/1002-137X-2026-53-1-216.pdf; https://doaj.org/toc/1002-137X; https://doaj.org/article/70cd6768a24344848a39d3848b97ea11 |
| DOI: |
10.11896/jsjkx.250300045 |
| Availability: |
https://doi.org/10.11896/jsjkx.250300045; https://doaj.org/article/70cd6768a24344848a39d3848b97ea11 |
| Accession Number: |
edsbas.5AF498B7 |
| Database: |
BASE |