An experimental comparison of visual SLAM systems

摘要

Monocular vSLAM (Visual Simultaneous Localisation and Mapping) is a navigation technique in which autonomous robots use visual information from a single camera to model a map of unknown environments while simultaneously localising themselves on that map. This paper explores prevalent algorithms’ performance in a highly dynamic use case on a quadcopter where other sensors’ data (viz. IMU, optical flow data) is not accessible. Three monocular vSLAM methods prevalent in the research community, Parallel Tracking and Mapping, ORB-SLAM and its enhancement ORB-SLAM2 (with 4585, 3936 and 2538 citations respectively) performed poorly in our use case. Better results were obtained using a Convolutional Neural Network based approach designed for ground-based vehicles. This paper presents an experimental comparison of these algorithms on metrics such as point cloud generation and localisation. Experiments were conducted physically and in the Gazebo simulator on the DJI Tello, a nano-quadcopter with an onboard camera, where we chose not to use the IMU data provided by the API.