High-precision visual navigation and localization method

Abstract

The paper studies high-precision visual navigation and localization technology which has avoided the impact of dynamic objects. The dynamic feature points in the scene are eliminated by the dynamic object detection algorithm based on the YOLACT instance segmentation and geometric epipolar constraint, and the algorithm is integrated into SLAM to achieve the improved three-dimensional location of the dynamic scene, based on the TUM RGB-D data set, the localization accuracy and the running speed of algorithm are tested respectively. The new algorithm is compared with traditional PL-SLAM and ORB-SLAM algorithms. The test results show that the new algorithm average running time on the three data set is less 30.229s than the traditional PL-SLAM algorithm. Compared with the traditional PL-SLAM algorithm and ORB-SLAM algorithm, the root mean square error of absolute trajectory is reduced by 0.084m and 0.130m respectively.