基于深度学习的动态环境视觉里程计研究  

Visual Odometry of Dynamic Environment Based on Deep Learning

作　　者：崔立志[1,2] 杨啸乾 杨艺 CUI Lizhi;YANG Xiaoqian;YANG Yi(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo 454003,China;Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment,Jiaozuo 454003,China)

机构地区：[1]河南理工大学电气工程与自动化学院,焦作454003 [2]河南省煤矿装备智能检测与控制重点实验室,焦作454003

出　　处：《空间控制技术与应用》2023年第6期58-67,共10页Aerospace Control and Application

基　　金：国家自然科学基金-联合基金项目(U1804147)。

摘　　要：本文提出了一种基于深度学习的动态场景视觉里程计方法。使用轻量级Ghost模块与目标检测网络YOLOv5s结合构建C3Ghost模块,引入坐标注意力机制(coordinate attention, CA),在提高网络检测速度的同时保证检测准确性。并将其与运动一致性算法结合,剔除动态特征点,仅利用静态特征点进行位姿估计。实验结果表明,与传统的ORB-SLAM3(orient FAST and rotated BRIEF-simultaneous localization and mapping 3)算法相比,在慕尼黑工业大学(technical university of Munich, TUM)RGB-D(RGB-depth)高动态数据集上绝对轨迹误差(absolute trajectory error, ATE)和相对位姿误差(relative pose error, RPE)平均有了90%以上的改善。相较于先进的同时定位与地图构建SLAM算法,也有相对提升。因此,该算法有效提升了视觉SLAM在动态环境下的稳定性和鲁棒性。This paper proposes a dynamic scene visual odometry method based on deep learning.The C3Ghost module is built using the lightweight Ghost module combined with the target detection network YOLOv5s,and the CA(coordinate attention mechanism)is introduced to improve the network detection speed while ensuring detection accuracy.It is combined with the motion consistency algorithm to eliminate dynamic feature points and only use static feature points for pose estimation.Experimental results show that compared with the traditional ORB-SLAM3(orient FAST and rotated BRIEF-simultaneous localization and mapping 3)algorithm,the ATE(absolute trajectory error)and RPE(relative pose error)on the TUM(technical university of Munich)RGB-D(RGB-depth)high dynamic data set has improved by more than 90%on average.Compared with the advanced SLAM algorithm,it is also relatively improved.Therefore,this algorithm effectively improves the stability and robustness of visual SLAM in dynamic environments.

关 键 词：视觉里程计 目标检测 注意力机制 轻量级 运动一致性 

分 类 号：V412.4[航空宇航科学与技术—航空宇航推进理论与工程]