动态场景下基于视觉特征的SLAM方法  

SLAM Method Based on Visual Features in Dynamic Scene

作　　者：张金凤 石朝侠[1] 王燕清 ZHANG Jinfeng;SHI Chaoxia;WANG Yanqing(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;School of Information Engineering,Nanjing Xiaozhuang University,Nanjing 211171,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]南京晓庄学院信息工程学院,南京211171

出　　处：《计算机工程》2020年第10期95-102,共8页Computer Engineering

基　　金：国家自然科学基金面上项目(61371040)。

摘　　要：同时定位与地图构建(SLAM)作为机器人领域的研究热点,近年来取得了快速发展,但多数SLAM方法未考虑应用场景中的动态或可移动目标。针对该问题,提出一种适用于动态场景的SLAM方法。将基于深度学习的目标检测算法引入到经典ORB_SLAM2方法中,将特征点分为潜在动态特征点和非潜在动态特征点,基于非潜在动态特征点计算运动模型,筛选出应用场景中的静态特征点并实现位姿跟踪,利用非潜在动态特征点中的静态特征点进行地图构建。KITTI和TUM数据集上的实验结果表明,与ORB_SLAM2系统相比,该方法能够提高跟踪轨迹精度与地图的适用性。As a research hotspot in the field of robotics,Simultaneous Localization and Mapping(SLAM)has made great progress in recent years,but few SLAM methods take dynamic or movable targets in the application scene into account.To handle the problem,this paper proposes a SLAM method which introduces the deep learning-based object detection algorithm into the classic ORB_SLAM2 method to make it more suitable for dynamic scene.The feature points are divided into dynamic feature points and potential dynamic feature points.The motion model is calculated based on dynamic feature points,which is used to select the static feature points in the application scene for pose tracking,and select static feature points in the dynamic feature points for map construction.Experimental results on KITTI and TUM datasets show that compared with the ORB_SLAM2 system,the proposed method improves the tracking accuracy and the application performance of the map.

关 键 词：同时定位与地图构建 局部特征 动态场景 深度学习 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]