PO-SLAM: A Novel Monocular Visual SLAM with Points and Objects

This paper proposes a semantic information-based real-time, robust, and low scale drift monocular Simultaneously Localization and Mapping (mono-SemSLAM) method. In this study, semantic information of the environment is intensely employed to achieve performance improvement of the front-end and back-end modules. Specifically, the proposed front-end extracts object semantic information from the image inputs by utilizing the objects detection network in real-time. The prior object size information used in the mono-SLAM helps to eliminate the scale uncertainty during the initialization. The proposed bundle adjustment optimizes landmark points, camera poses, and map scale to correct scale drift and hence improve the accuracy of the mono visual SLAM system. The proposed method is tested and evaluated in the KITTI odometry datasets. The experiment results show that our method can achieve better performance than the state of art mono visual SLAM system-ORB-SLAM2 in real-time mapping accuracy and scale drift elimination.