近年来,视觉实时定位与建图技术(visual simultaneous localization and mapping,VSLAM)技术取得了快速发展.然而,大多数传统的VSLAM系统存在鲁棒性较差的问题而基于深度学习的VSLAM存在精度较低等问题.为了提高VSLAM系统的性能,提出了一种基于无监督深度学习的单目VSLAM.该方法首先设计深度估计网络完成对图像的深度估计.然后,设计位姿估计网络进行相机的位姿估计.最后,使用合理的损失函数保证网络在迭代过程中的有效收敛.在KITTI数据集上验证了该系统的性能.实验结果表明,与SfMLearner相比,绝对轨迹误差(absolute trajectory error,ATE)降低大约50％.与传统的VSLAM系统相比,绝对位姿误差(absolute pose error,APE)的平移部分误差也明显下降且鲁棒性得到提升.
Monocular Visual Odometry Based on Unsupervised Deep Learning
VSLAM(visual simultaneous localization and mapping)technology makes rapid progress in recent years.However,most traditional VSLAM systems have poor robustness,while VSLAM based on deep learning have low accuracy.In order to improve the performance of VSLAM systems,an unsupervised deep learning based monocular VSLAM was proposed.The depth estimation net-work was designed to complete the depth estimation of the image and the pose estimation network was applied to estimate the pose of the camera.Finally,a reasonable loss function was used to ensure the effective convergence of the network during the iterative process.The performance of the system is verified on KITTI data set.The results show that ATE(absolute trajectory error)is reduced by 50％compared to SfMLearner.Compared with traditional VSLAM system,the translation part of APE(absolute pose error)is also signifi-cantly reduced and the robustness is also improved.
SLAM roboticsdepth estimation networkpose estimation networkunsupervised learning
NETL
NSTL
万方数据
