动态场景下基于图像掩模技术的双目SLAM算法  

Binocular SLAM Algorithm Based on Image Mask Technology in Dynamic Scenes

作　　者：王慧颖 吴琦鸣 王兆强 WANG Huiying;WU Qiming;WANG Zhaoqiang(Department of Foundation,China Fire and Rescue College,Bejing 102202,China;Department of Foundation,Rocket Force Engineering Unirersity,Xi'an Shaanxi 710025,China)

机构地区：[1]中国消防救援学院基础部,北京102202 [2]火箭军工程大学基础部,陕西西安710025

出　　处：《传感技术学报》2021年第12期1656-1662,共7页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(61603398,61873273)。

摘　　要：视觉SLAM技术在动态场景中存在定位漂移和创建地图扭曲的情况,本研究提出一种基于图像掩模技术的双目SLAM算法,运用视觉几何约束法在动态场景中得出动态特征点,将检测出的运动物体区域作为图像掩模。在ORB-SLAM算法中引入动态区域掩模分割和剔除动态区域的功能,可以保证相机的定位和构图不受动态场景中的运动物体的影响。根据动态特征点和分割结果在动态场景中用矩形框标识出动态区域掩模,剔除特征匹配中动态特征点,从而可以降低运动物体对SLAM算法精度的影响。经实验得出,改进算法较双目ORB-SLAM算法在动态场景中的定位精度提高76.1%,其动态场景构图效果改善明显,算法处理的平均速度可以达到实时视觉定位和构图的需求,平均可达4.7 frame/s。As visual SLAM exists positioning drift and composition in dynamic scene distortions,binocular vision SLAM algorithm is proposed based on image mask technology.Dynamic characteristic points are concluded by applying the method of visual geometric constraints in the scene,which can detect moving objects in dynamic area and call it as image masks.The function of the dynamic region mask segmentation and removing dynamic region is introduced into ORB-SLAM algorithm,which ensures that camera positioning is not affected by the moving object and composition.According to the dynamic feature points and segmentation results,the dynamic region mask is identified by a rectangular frame,and the dynamic feature points are removed to reduce the influence of moving targets on the accuracy of SLAM.Experimental results show that the improved algorithm improves positioning accuracy by 76.1%in comparison with binocular ORB-SLAM algorithm,and the composition effect is significantly improved.The average speed of the improved algorithm reaches real-time,and the average speed can reach 4.7 frames/second.

关 键 词：同步定位与地图创建(SLAM) 双目相机 动态场景 动态区域掩模 区域分割 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]