Abstract: SLAM(Simultaneous Localization and Mapping) technology is a camera pose estimation based on static scene features. It is susceptible to dynamic objects in the process of feature calculation and matching at its front end. Therefore a method of instance segmentation combined with multi-view geometric constraints is proposed to improve the front-end feature processing of visual SLAM and eliminate the interference of dynamic information. Specifically, in the front end of the ORB-SLAM3(Oriented FAST and Rotated BRIEF-Simultaneous Localization and Mapping3) framework, the YOLACT + + (You Only Look At CoefficienTs + +) instance segmentation thread is paralleled, and the segmented results are used to supplement the multi-view geometric constraint method testing the dynamic consistency of feature points. The EfficientNetV2 network is used to replace the original backbone network of YOLACT + +, and the TensorRT is used to quantify the instance segmentation model to reduce the front-end computing pressure of the algorithm. The test of TUM data set shows that the positioning accuracy of the proposed algorithm in high dynamic environment is 80. 6% higher than that of ORB-SLAM3 algorithm. 

Key words: semantic simultaneous localization and mapping(SLAM), YOLACT++ segmentation algorithm, multi-view geometric constraints, dynamic scene