室内动态场景下基于稀疏光流与实例分割的视觉SLAM算法

Visual SLAM algorithm based on sparse optical flow and instance segmentation for indoor dynamic scenes

针对传统视觉SLAM算法在动态场景下难以兼顾算法定位精度与实时性的问题,提出一种基于改进Tracking线程的动态视觉SLAM算法。针对算法定位精度,以ORB-SLAM2算法为基础,在算法框架中融入Lucas-Kanade稀疏光流和Mask R-CNN图像分割网络,提高算法对潜在动态对象的感知能力。针对算法的实时性,利用Mask R-CNN图像分割网络设计异常特征点二次剔除算法,降低图像分割网络对Tracking线程耗时的影响。基于TUM与Bonn公开数据集验证算法的有效性。结果表明,改进算法的定位误差相比ORB-SLAM2、ReFusion、Dyna-SLAM算法最高降低98.62%、96.18%和91.82%,跟踪线程耗时相比ReFusion、Dyna-SLAM算法最高降低83.32%、88.72%,改进算法可同时兼顾算法的定位精度与实时性。

A dynamic visual SLAM algorithm is proposed based on improved tracking algorithm to cope with the problems that the traditional visual SLAM is difficult to give consideration to positioning accuracy and real-time performance.Aiming at the positioning accuracy of the algorithm,Lucas-Kanade sparse optical flow and Mask R-CNN image segmentation network were incorporated into the ORBSLAM2 algorithm to improve the perception ability of the algorithm to potential dynamic objects.Aiming at the real-time performance of the algorithm,a secondary elimination algorithm of abnormal feature points was designed by Mask R-CNN image segmentation network,to reduce the impact of image segmentation network on the time-consuming of the Tracking thread.Experiments were carried out on TUM dataset and Bonn dynamic dataset to verify the effectiveness of the improved algorithm.The results show that compared with ORB-SLAM2,ReFusion and Dyna-SLAM algorithms,the positioning error of the improved algorithm can be reduced by 98.62%,96.18%and 91.82%at the highest.Compared with ReFusion and Dyna-SLAM algorithms,the tracking thread time can be reduced by 83.32%and 88.72%at the highest.The improved algorithm can take into account the positioning accuracy and real-time performance of the algorithm.