室内动态环境语义VSLAM应用研究

Research on semantic VSLAM application in indoor dynamic environment

针对VSLAM算法在室内动态场景下精度低、鲁棒性差的问题,结合深度学习和多视图几何的方法,首先利用语义分割网络对RGB-D视频流图像进行处理,建立动态对象的先验语义信息,然后针对无先验信息的移动物体进行多视图几何处理,最后使用静态特征点进行特征匹配和位姿估计。在慕尼黑工业大学开源的TUM公开数据下进行实验验证,结果显示,在室内动态环境下,该系统相较于ORB-SLAM2和ORB-SLAM3能够更接近真实轨迹,且轨迹的绝对位姿误差显著降低;在高动态环境下,相较于ORB系列算法,平均提升了79.1625%;在低动态环境下,平均提升了12.4125%。

In response to the low accuracy and poor robustness of the VSLAM algorithm in indoor dynamic scenes,a combination of deep learning and multi view geometry methods is used.Firstly,a semantic segmentation network is used to process RGB-D video stream images,establish prior semantic information of dynamic objects,then perform multi view geometric processing on moving objects without prior information,and finally use static feature points for feature matching and pose estimation.Experimental verification was conducted on TUM open source data at the Technical University of Munich,and the results showed that in indoor dynamic environments,the system was closer to the real trajectory compared to ORB-SLAM2 and ORB-SLAM3,and the absolute pose error of the trajectory was significantly reduced.In high dynamic environments,compared to the ORB series algorithms,the average improvement is 79.162%.In low dynamic environments,an average improvement of 12.4125%was achieved.