Fusion of visual odometry and inertial data for enhanced real-time egomotion estimation

摘要

This paper presents a real-time motion estimation module for ground vehicles based on the fusion of monocular visual odometry and low-cost inertial measurement unit data. The system features a novel algorithmic scheme enabling accurate and robust scale estimation and odometry at high speeds. Results of multiple performance characterization experiments (on rough terrain at speeds up to 20 mph and smooth roadways at speeds of up to 75 mph) are presented. The prototype system demonstrates high levels of precision (relative distance error less than 1%, and less than 0.5% on paved roads, yaw drift rate ~2 degrees per km) in multiple configurations, including various optics and vehicles. Performance limitations, including those specific to monocular vision, are analyzed and directions for further improvements are outlined.