动态场景感知下的移动机器人视觉定位与建图  

Research on Improved Visual Positioning and Mapping under Dynamic Scene Detection

作　　者：隆良梁 魏小源 LONG Liangliang;WEI Xiaoyuan(Mianyang Polytechnic,Mianyang Sichuan 621000,China;College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou Gansu 730050,China;Key Laboratory of Gansu Advanced Control for Industrial Processes,Lanzhou Gansu 730050,China)

机构地区：[1]绵阳职业技术学院,四川绵阳621000 [2]兰州理工大学电气工程与信息工程学院,甘肃兰州730050 [3]甘肃省工业过程先进控制重点实验室,甘肃兰州730050

出　　处：《机床与液压》2024年第21期79-86,共8页Machine Tool & Hydraulics

基　　金：国家自然科学基金青年科学基金项目(62101228);甘肃省科技重点研发项目(20YF3GA018)。

摘　　要：移动机器人可替代或辅助人类的生产和生活,并深入复杂或危险的环境,需随时感知周围场景。传统的ORB-SLAM2方法在动态场景中定位不准确,由此提出一种动态场景感知下的移动机器人视觉定位与建图方法。SLAM方法通过光学传感器采集环境信息完成建图,以ORB-SLAM2作为基本框架,在其跟踪线程中添加一个动态目标检测模块,该模块采用YOLOv4网络检测目标,使用的CSPDarknet53结构在减轻骨干网络权重的同时保持检测准确性;采用改进四叉树算法提取特征,并采用改进型形状上下文的图像匹配方法完成特征匹配。在TUM RGB-D数据集上进行实验,所提算法在walking_xyz场景序列的RMSE相对ORB-SLAM2算法增加97.6%,walking_rpy场景序列的RTE和RRE分别改进97.1%和96%。比较所提算法与ORB-SLAM2算法在高、低动态场景中的客观指标和估计轨迹误差,所提算法的RMSE、均值、中间值以及性能提升的改进百分比较优,所估计出的轨迹能够更好地拟合真实情况。Mobile robots can replace or assist human production and life,and go deep into complex or dangerous environments,and need to feel the surrounding scene at any time.The traditional ORB-SLAM2 algorithm is inaccurate in dynamic scenarios.A fast feature point extraction method based on dynamic detection was proposed to describe visual localization and mapping.In SLAM algorithm,environmental information was collected through optical sensors to complete the construction of the map.ORB-SLAM2 was used as the basic framework,and a dynamic target detection module was added to its tracking thread.The module adopted YOLOv4 network to detect the target and CSPDarknet53 structure was used to reduce the weight of the backbone network and maintain the detection accuracy.An improved quadtree algorithm was used to extract features,and an improved shape context image matching method was used to complete feature matching.The experiments were conducted on the TUM RGB-D dataset.Compared with the ORB-SLAM2 algorithm,the RMSE of sitting_xyz scene sequence is increased by 97.6%,and the RTE and RRE of walking_rpy scene sequence are improved by 97.1%and 96%,respectively.The objective indicators and estimated trajectory errors of the proposed method are compared with that of the ORB-SLAM2 algorithm in high/low dynamic scenarios.The RMSE,mean value,median value and performance improvement percentage of the proposed method are superior,and the estimated trajectory can better fit the real situation.

关 键 词：移动机器人 视觉定位与建图 动态场景 目标检测 图像匹配 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置] TN958.98[自动化与计算机技术—控制科学与工程]