摘要
针对暗光环境下特征丢失影响视觉同步定位与地图构建(SLAM)精度的问题,提出一种深度可分离U型网络(DSCU-net)的图像增强方法:参考编码解码结构与跳跃连接机制,构建逐像素变换曲线估计网络,并引入深度可分离卷积以减少网络参数量;然后在公开数据集上进行图像增强算法性能测试,并使用开源SLAM算法验证DSCU-net对定位精度的影响。结果表明,该方法能有效提升图像照明度,降低暗光条件下的定位误差,最小误差可降至4.9 cm;综合考虑增强图像质量和计算效率,提出的方法具有优越的暗光增强性能和网络轻量化特点,能有效提高暗光环境下视觉SLAM的定位精度。
Aiming at the problem that feature loss affects the accuracy of visual simultaneous localization and mapping(SLAM)in low-light environments,the paper proposed a deep separable U-net(DSCU-net)for image enhancement:inspired by an encoder-decoder structure with skip connections,a per-pixel transformation curve estimation network was built,and the deep separable convolutions were incorporated to reduce network parameters;then,the performance of image enhancement algorithm was tested on public datasets,and the impact of DSCU-net on localization accuracy was validated by using open-source SLAM algorithms.Results showed that the proposed method could effectively improve the brightness of images and reduce the positioning errors under low-light conditions,achieving a minimal error reduction to 4.9 cm;in general,balancing enhanced image quality and computational efficiency,the proposed method would excel in low-light enhancement and lightweight network characteristics,and could effectively enhancing the localization precision of visual SLAM under dark environments.
作者
石秋婷
程玉
陈帅
吴奕雯
陈垚杰
SHI Qiuting;CHENG Yu;CHEN Shuai;WU Yiwen;CHEN Yaojie(School of Automation,Nanjing University of Science and Technology,Nanjing 210094,China;School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《导航定位学报》
北大核心
2025年第1期106-112,共7页
Journal of Navigation and Positioning
关键词
暗光
视觉定位
同步定位与地图构建(SLAM)
图像增强
深度可分离卷积
轻量化
low-light
visual localization
simultaneous localization and mapping(SLAM)
image enhancement
depth-wise separable convolutions
lightweight
作者简介
第一作者:石秋婷(1998—),女,福建漳州人,硕士研究生,研究方向为视觉与惯性组合导航。
