A visual localization technology in low illumination scenes
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
维普
万方数据
针对在低照度环境中由于光照不足或光照不均导致的图像噪声过大、特征提取不均匀等问题,提出了一种低照度场景下的单目视觉定位技术.首先,利用微光传感器采集低照度图像信息,对图像噪声设计了一种基于深度学习的图像去噪网络,利用该网络进行图像噪声处理.然后,利用四叉树改进特征均匀提取策略以提高特征跟踪效果,采用对极几何、三角测量等技术估计图像帧间位姿.最后,构建视觉重投影误差方程,利用光束平差法进行位姿估计和优化.实验结果表明,所提定位技术在光照强度为0.01 lx的低照度环境中,轨迹闭环情况下的平均定位均方根误差小于1.47 m,轨迹无闭环情况下的平均定位均方根误差小于4.26 m.
In order to solve the problems of excessive image noise and uneven feature extraction in low-light environment caused by insufficient or uneven illumination,a monocular visual localization technology in low-light scenes is proposed.First of all,the low-light sensor is used to collect low-light image information.Aiming at the problem of image noise,an image denoising network based on deep learning is designed,and the network is used to process image noise.Then,the quadtree is used to improve the feature uniform extraction strategy,and the feature tracking effect is improved.The image inter-frame pose is estimated by using epipolar geometry,triangulation and other techniques.Finally,the visual reprojection error equation is constructed,and the bundle adjustment method is used for pose estimation and optimization.The experimental results show that the average location root mean square error of the proposed technology is less than 1.47 m when the trajectory has a closed loop,and the average location root mean square error is less than 4.26 m when the trajectory has no closed loop in the low illumination environment of the light intensity of 0.01 lx.
维普
万方数据
