Vision-Based Trajectory Tracking Approach for Mobile Platforms in 3D World Using 2D Image Space

Abstract

Vision-based target following using a camera system mounted on a mobile platform has been a challenging problem. A scenario is assumed in which the platform/camera system is required to have a desired trajectory for relative position and orientation (pose) with respect to a target object. It is assumed that the actual pose of mobile platform with respect to the target is not measured by a Global Positioning System (GPS) and/or an Inertial Measurement Unit (IMU). For trajectory tracking feedback control, the error in the relative pose of the mobile platform with respect to the target needs to be computed. In the absence of GPS/IMU signals, the error in relative pose must be calculated using a vision-based approach. In this paper, we introduce a fast alternative vision-based approach for real-time calculation of the error in the relative pose of the mobile platform and the target. The proposed vision-based tracking approach is called PIVOT: Positioning and Orienting Using Vision-Based Object Tracking. The PIVOT system calculates the pose errors of the mobile platform in the 3D world based on a 2D image space. The only information required is “the desired 3D pose” and the coordinates of selected feature points on the target in order to track properly. The PIVOT forms a desired target image, compares it with the current target image and outputs the required 3D translation and rotation of the platform/camera to correct the image error. The required 3D translation and rotation of the platform/camera to correct the image error are fed to a feedback controller to drive the mobile platform in the direction that corrects the image error. When the image error is vanished, the mobile platform is moving on its desired trajectory. We have performed a set of experiments with the proposed PIVOT approach to show the effectiveness of the theoretical framework. According to the simulation results, PIVOT provides accurate pose errors for all test cases. The formulation of the approach is general, such that it can be applied to mobile platforms that move in 3D as well as 2D. Our first simulated and experimental tests will be on a mobile robot.