Spatio-Temporal Fusion-based Monocular 3D Lane Detection


Yin Wang (JiLin University),* Qiuyi Guo (SenseTime Group Limited), Peiwen Lin (SenseTime Group Limited), Guangliang Cheng (Sensetime Group Limited), Jian Wu (Jilin university)
The 33rd British Machine Vision Conference

Abstract

The monocular 3D lane detection (Lane3D) methods are increasingly proposed to address the issue of inaccurate bird-eye-view (BEV) results in various complex scenarios (e.g. up and downhills, bumps). However, there are a few restrictions on existing Lane3D methods. Primarily, only single-frame input is considered, which leads to poor results in no visual cues scenarios (e.g. obscured by surrounding vehicles). The other is that they rely on the camera pose to the road surface to translate to the road coordinate system. To address these issues and better exploit the spatio-temporal continuity of the lanes, we propose a novel Spatial-Temporal Lane3D model abbreviated as STLane3D. First, we propose a novel multi-frame pre-alignment layer under BEV, which uniformly projects features from different frames onto the same ROI region. Afterward, we propose a plug- and-play spatio-temporal attention module and a new 3DLane IOULoss. Experiments on the ONCE and OpenLane datasets demonstrate that our single-frame model, independent of camera extrinsic, can achieve close detection accuracy compared to the current state- of-the-art. And our multi-frame model improves the F1 score by 3.5% compared to the single-frame model on the ONCE dataset, which demonstrates the effectiveness of the multi-frame fusions strategy. Moreover, with multi-frame information, our model achieves satisfying performance in complex scenes lacking enough visual information and meets the real-time requirements on autonomous vehicles.

Video



Citation

@inproceedings{Wang_2022_BMVC,
author    = {Yin Wang and Qiuyi Guo and Peiwen Lin and Guangliang Cheng and Jian Wu},
title     = {Spatio-Temporal Fusion-based Monocular 3D Lane Detection},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {https://bmvc2022.mpi-inf.mpg.de/0314.pdf}
}


Copyright © 2022 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection