The application of Kalman filter in visual odometry for eliminating direction drift

The paper describes a visual odometry system based on stereovision camera and Kalman filter. A common problem of odometry is that a next pose is estimated on the base of a previous one. Once a location is measured incorrectly, the error propagates on next locations. Inaccuracies quickly accumulate, which at the end leads to significant errors. The system is most sensitive to direction estimation, as it exhibits small rotation drifts, that quickly sum up to a large error. To limit this effect, the algorithm selects a distant point which is used as a landmark and serves as an angle of reference. Due to occlusions (e.g. humans, cars), the point of reference is not always visible or cannot be singled out (e.g. in park alleys with tree branches hanging down). Then, the system uses differential rotation measurement. The differential and landmark angle estimates, together with their uncertainties, are connected by Kalman filter to get the optimal direction.