An AR Projection Improvement Strategy via the Integration of Target Detection and ORB-SLAM2

Abstract

Visual simultaneous localization and mapping (vSLAM) algorithms are mainstream technical methods for markerless augmented reality using monocular cameras. However, most vSLAM algorithms are incompetent to project virtual object on a specified plane, especially when rigour precision is required. This is because they fail to distinguish the mappoints of interest from other normal ones. In this work we propose a new SLAM system which integrates target detection algorithm into conventional ORB-SLAM2 so that mappoints of interest can be regarded as variables and hence be optimized continuously. Specifically, the proposed system adds a new class member called Target in the map of ORB-SLAM2 so that the system can detect the target during operation with Oriented Fast and Rotated Brief (ORB) features and distinguish the Target Mappoints from others. Compared to conventional ORB-SLAM2, proposed system needs to take on two more tasks: management of Target Mappoints and updating the projection matrix in all three treads. In this work we use the Lena picture as the target plane we want to project visual objects on and the test results demonstrate that our system can perform projection more accurately.