Reducing drift using visual information in a multi-robot scheme

摘要

In this work, we use computer vision to improve odometry result of multiple robots with shared information amongst neighboring region. We use a stereo camera to captures images. Each robot handles visual odometry to localize itself using stereo images. For every frame that is captured, there are errors in calculations, and these errors do accumulate and cause drift in odometry results. To solve this issue, we are using the information from neighboring robot. Our main assumption is that robots do have a shared field of view more than once, but they do not need to face towards the same direction at all times. Robot-to-robot measurements are computed at certain time intervals. Individual robots do not have the scale problem, on the other hand, we need to calculate scale for robot-to-robot measurements. Initially, we use monocular pairs to calculate the relative transformation between two robots. Then we refine rotation translation by using point clouds calculated from stereo pairs of each robot. Finally, we combine individual measurements and robot-to-robot measurements to utilize graph optimization tools to decrease drift. We have conducted experiments on both simulations and with a robot.