Long-term Accuracy of Camera and IMU Fusion-based Navigation Systems

Recently, several navigation approaches which fuse together visual and inertial information have been proposed. Each of these approaches claim significant advantages over navigation with an inertial measurement unit (IMU) only. In this paper, we propose a method for analytically evaluating the performance of two commonly implemented visual and inertial fusion based methods. Similar to the "Big O" notation used in algorithms analysis, we propose to analyze proposed navigation algorithms according to their "drift order" (D). The drift in location and attitude estimates of a navigation algorithm as time goes to infinity (its drift order) defines its performance independent of the specific amounts of noise in the sensors used for navigation, enabling a quantitative comparison between algorithms. We analyze the drift order of an IMU-only navigation solution and two common methods for fusion of inertial and visual information. Our analysis shows that SLAM-based methods for visual and inertial fusion are of a significantly lower ;drift order (D(t)) than IMU-only and epipolar-based fusion navigation approaches (D(t(3))). We present simulation results verifying our analysis.