由于视觉SLAM(Simultaneous Localization and Mapping)算法研究多建立于静态环境中，使得在动态环境下的应用造成较大定位偏移，极大降低了系统的稳定性。针对该问题，该文在原有视觉SLAM算法的基础上结合深度学习方法，对环境可能存在的动态目标进行特征点剔除，从而提升系统在动态环境下的鲁棒性。采用的视觉 SLAM 系统为 ORB-SLAM3，深度学习方法为YOLOv5 的实例分割算法，采用对目标模型mask轮廓内特征点的检测算法及多视角几何方法进行特征点剔除。首先利用并行通信，将SLAM系统获取到的帧数据传入YOLOv5 系统中进行可能为动态目标的分割，然后将其分割结果传回SLAM系统进行跟踪建图。同时改进词袋加载模型，提升加载速度，最终构建动态环境的稠密地图，具备可靠的实时性。通过在TUM数据集上的实验评估，该方法对比原SLAM框架及现阶段经典动态环境研究均有提升，其在保证平均帧率不降低的前提下精度较ORB-SLAM3 的RMSE平均提升近 89%。实验结果表明，对动态环境下的视觉SLAM算法有效改进，极大提升了系统的鲁棒性及稳定性。
Optimization of Visual SLAM Algorithm for Fusion Instance Segmentation in Dynamic Environment
Because the research of visual SLAM(Simultaneous Localization and Mapping)algorithm is mostly established in the static environment,the application in the dynamic environment causes large positioning offset,which greatly reduces the stability of the system.To solve this problem,based on the original visual SLAM algorithm,we combine the deep learning method to eliminate the feature points of the dynamic targets that may exist in the environment,so as to improve the robustness of the system in the dynamic environment.The visual SLAM system used is ORB-SLAM3,the deep learning method is the instance segmentation algorithm of YOLOv5,and the feature point detection algorithm in the contour of the target model mask and the multi-view geometric method are used to remove the feature points.Firstly,by using parallel communication,the frame data obtained by the SLAM system is sent to the YOLOv5 system for the detection of possible dynamic targets,and then the detection results are sent back to the SLAM system to add multi-view geometric dis-crimination for tracking and mapping.At the same time,the bag-of-words loading model is improved to improve the loading speed,and finally the dense map of the dynamic environment is constructed,which has reliable real-time performance.Through the experimental e-valuation on the TUM dataset,the proposed method has an improvement compared with the original SLAM framework and the current classic dynamic environment research.The accuracy of the proposed method is improved by nearly 89% on average compared with the RMSE of ORB-SLAM3 without reducing the average frame rate.The results show that the visual SLAM algorithm in dynamic environment is effectively improved,and the robustness and stability of the system are greatly improved.
dynamic environmentvisual SLAMfeature point detectioninstance segmentationdense mapping
万方数据
维普
