Improved Visual Odometry Based on SSD Algorithm in Dynamic Environment

Abstract

In recent years, Simultaneous Localization and Mapping (SLAM) have made great progress in the field of robots. However, the majority of SLAM algorithms are currently designed and processed in static environments. In the dynamic environment, due to the existence of moving objects, the SLAM system is easy to mismatching the pose estimation, which will affect its positioning accuracy and mapping accuracy. In this paper, the traditional Oriented FAST and Rotated Brief (ORB)-SLAM2 system is improved to deal with these problems. Firstly, the Single Shot MultiBox Detector (SSD)algorithm is used to detect the dynamic targets in the scene, and then the multi-layer optical flow is constructed to remove the dynamic feature points by optical flow tracking. Secondly, the static feature points in the scene are matched and the pose are calculated to solve the problems caused by a large number of matches, such as large system computation and low operating efficiency. Finally, testing on the TUM datasets shows that the absolute trajectory error and the relative pose error of the improved ORB-SLAM2 system are significantly reduced compared to the traditional ORB-SLAM2 system.