End-to-End Multi-Level Encoding Methods of Visual Data Compression for Robust Monocular Visual ORB-SLAM

Abstract

Simultaneous localization and mapping (SLAM) has been highly studied in the last decade. It allows the estimation of the camera pose of a mobile device and the creation of a map of the surrounding environment concurrently. Recently, Visual SLAM (VSLAM) has become the most widely used state-of-the-art technique to implement SLAM tasks due to its reduced cost, lower size, and affordability. However, the intensive computation of VSLAM systems does not fit in a wide range of limited resources and energy mobile devices. A possible solution is to split its functionality between mobile devices and the edge cloud. This solution showed the necessity for efficient visual data compression methods to be integrated within VSLAM systems. This work proposes a multi-level encoding method for visual data frame compression integrated within the monocular Oriented FAST and Rotated BRIEF-SLAM (ORB-SLAM) system. The performance results of the proposed system are compared to corresponding ORB-SLAM systems adopting the most popular classical still image compression standards; the Joint Photographic Experts Group (JPEG) and the advanced version, the JPEG 2000, in terms of reconstruction quality, robot’s trajectory estimation, and computational complexity.