摘要
针对移动机器人位姿估计过程中存在累积误差导致位姿估计精度差的问题,提出一种基于深度学习方法的视觉里程计。首先,设计卷积神经网络,通过逐层优化卷积核的大小,提取图像序列更多的细节特征;然后,利用自适应存储网络记录历史位姿信息,通过双向长短期记忆网络预测未来的位姿信息,将历史信息和未来信息同时作用于当前时刻的位姿输出,降低累积误差对位姿估计精度的影响;最后,在KITTI和TUM数据集上进行仿真实验,实验结果表明,相比于其他视觉里程计算法,所提算法的位姿估计精度、绝对轨迹误差和相对位姿误差均有很大改善。
In order to solve the problem of pose estimation accuracy degradation caused by error accumulation in mobile robot localization,a deep learning-based visual odometry method is proposed.Firstly,a convolutional neural network(CNN)is designed to extract more detailed features of the image sequences by optimizing the size of the convolutional kernel layer by layer.Then,an adaptive memory network records historical poses,while a bi-directional long short-term memory(Bi-LSTM)predicts future poses.By fusing both past and future information,the method reduces error accumulation in pose estimation.Finally,experiments on KITTI and TUM datasets show the method outperforms existing approaches in pose accuracy,absolute and relative trajectory error.
作者
刘海涛
戴娟
朱胜涛
李剑锋
LIU Haitao;DAI Juan;ZHU Shengtao;LI Jianfeng(Beijing Key Laboratory of High Dynamic Navigation Technology,Beijing Information Science&Technology University,Beijing 100192,China;Key Laboratory of Modern Measurement&Control Technology,Ministry of Education,Beijing Information Science&Technology University,Beijing 100192,China;College of Automation,Beijing Information Science&Technology University,Beijing 100192,China)
出处
《控制工程》
北大核心
2025年第9期1611-1618,共8页
Control Engineering of China
基金
国家自然科学基金资助项目(61703040,61603047)
北京信息科技大学师资补充与支持计划(2019-2021)(50290 11103)
北京信息科技大学科研水平提高重点研究培育项目(2121YJPY221)
高动态导航技术北京市重点实验室基金资助项目(HDN2019001)。
关键词
视觉里程计
深度学习
相机位姿估计
卷积神经网络
双向长短期记忆网络
自适应存储
Visual odometry
deep learning
camera pose estimation
convolutional neural networks(CNN)
bi-directional long short-term memory(Bi-LSTM)network
adaptive memory
