Minimizing Odometry Drift by Vanishing Direction References

Robotic systems that operate in environments without access to global references like GNSS normally estimate their position by odometry. The incremental nature of this measurement is subject to drift that accumulates to large errors in the position estimate. We minimize this drift by exploiting scene knowledge in form of vanishing directions. These provide a scene referenced orientation measurement which we apply in an orientation filter to correct the estimated attitude and in turn improve the position estimation. The experiments show that the resulting accuracy is comparable to a referenced inertial measurement unit and that our method can thus augment or replace such sensors.