基于单目视觉的动态环境同步定位与多地图构建算法

Simultaneous Localization and Multi-mapping Algorithm in Dynamic Environment Based on Monocular Vision

针对传统V-SLAM算法是在假设场景刚性不变的条件下进行建图,导致无法实现动态环境建图,以及传统算法无法克服因环境特征不明显或机器人被“绑架劫持”而导致场景跟丢的问题,提出一种在动态环境下同步定位与多地图构建(DE-SLAMM)算法。该算法首先引入一种多地图构建思想,当跟踪失败时会自适应生成一个新的局部地图,并在回环时将该地图与之前地图融合,解决算法跟丢后无法建图的问题。其次,结合深度学习和多视图几何技术实现对环境中的动态物体进行实时检测,并利用多帧融合技术对动态对象遮挡的部分进行背景修复,有效解决动态环境下跟踪建图问题。最后将该算法应用于实际场景进行测试,结果表明,相比经典的V-SLAM算法(ORB-SLAM2、ORBSLAMM和DynaSLAM),当发生跟踪丢失时,本文算法在很短时间内快速重建地图并实现继续跟踪和新地图融合,而ORB-SLAM2和DynaSLAM跟丢后进入重定位模式,无法继续建图;ORBSLAMM跟丢后虽然可以继续建图,但其建立的地图不能实现多地图融合,无法构建整体地图;进一步通过动态环境测试实验发现,只有本文算法可实现所有动态目标(先验和移动目标)的实时检测及背景修复,DynaSLAM只能实现先验目标检测,而其它两种算法无法实现动态环境下目标检测和建图。

In view of the traditional V-SLAM algorithm,the map is created under the assumption that the rigidity of the scene remains unchanged,leading to the inability to achieve dynamic environment mapping,and traditional algorithms cannot overcome the problem of scenes being lost due to unobvious environmental features or robots being“kidnapped and hijacked”.An algorithm for simultaneous localization and multi-mapping in a dynamic environment was proposed.Firstly,the algorithm introduced a multi-mapping idea,when tracking failed,a new local map would be adaptively generated,and the map would be merged with the previous map during the loop,to solve the problem that the algorithm cannot be built after the algorithm was lost.Secondly,deep learning and multi-view geometry technology were combined to realize real-time detection of dynamic objects in the environment,and multi-frame fusion technology was used to repair the background of the parts occluded by dynamic objects,effectively solving the problem of tracking and mapping in a dynamic environment.Finally,the algorithm was applied to actual scenes for testing.The results showed that compared with the classic V-SLAM(ORB-SLAM2,ORBSLAMM and DynaSLAM)algorithm,when tracking loss occurred,the proposed algorithm can quickly rebuild the map in a short time.And to realize the continuous tracking and the new map integration,ORB-SLAM2 and DynaSLAM would enter the relocation mode after being lost,and cannot continue to build.Although ORBSLAMM can continue to build maps after being lost,the built maps cannot achieve multi-mapping integration and cannot build an overall map;further through dynamic environment test experiments,it was found that only the algorithm can achieve all dynamic goals(a priori and moving goals).For real-time detection and background restoration,DynaSLAM can only achieve a priori target detection,while the other two algorithms cannot achieve target detection and mapping in a dynamic environment.