Closing the loop in appearance-guided omnidirectional visual odometry by using vocabulary trees