In this theses the wheel encoder based odometry was examined. In the Introduction the related works and method were written and a simulation of vehicle model based localization was presented, which shows that a well-calibrated vehicle model is required to ensure the highly accurate pose estimation. The dissertation stars with the theoretical background of Car-like robots, sensors in navigation, vehicle models and the methods of state and parameter estimation.
The second chapter starts with a simulation data generation using the Carmaker software. GPS position, magnetic heading and wheel impulse signals were used in a Kalman-filter to state fusion. Due to the necessity of parameter estimation the two-wheel vehicle model was used. The effect of the model and measurement covariances were examined through the simulation data, because ground truth values are available. The wheel circumference parameters are identified by two different methods. The well-known Augmented Kalman-filter handles the parameters as state variables. The results show that the setting of the optimal covariances is difficult, because using high values of model covariances provided similar estimated parameters as the setting of high measurement covariance values. Furthermore can not be realize a fix value of the wheel circumferences, because the values change in every timestamp, thus the actual uncertainty of measurements could cause offset too. After that, a Least Square based estimation was develop. It works in an iterative loop with two layers. The first is a state fusion mentioned before, which provides prior estimated pose data. The aim of second layer is to determine the optimal circumference, which provides the best fit of the preliminary estimated pose to the measurements. The appropriate values of covariances were specified easier in this method. The estimation was executed for 10 different generated noise. The average error of the parameters are only 1.2712 mm, which is 0.0647% of the true value and it is about the half of the result of the other method. However it also needs high manual tuning to determine the optimal covariances. To eliminate this, the first layer was complemented by a varying covariance solution and the odometry errors were used to find the optimal wheel circumferences. This provided 0.8710 mm average circumference error, which only 0.0444%. However the main advantage is that there is only one tuneable parameter in the filter, which responsible for the speed of convergence and the value of this is in a certain range. The fit of the odometry using the optimal parameters is close to perfect to the reference path. These simulation results show that the developed methods are feasible for model identification.
In the Chapter 3 and Chapter 4 the methods were presented on real data, which was recorded from a racecar, which is equipped with GPS, magnetometer, camera and LiDAR too. The landmark outputs of a perception algorithm were used in projection method. The prior displacement of the car between two timestamp was calculated using the odometry model and the result was suitable for data association and fault detection. The output of the projection algorithm is the optimal velocities and sideslip of the vehicle between two timestamp. The method is the general grid search, because the run-time disadvantage is unimportant for us. The average projection error below than 20 mm, and it was examined in detail that the estimated path is certainly more real than the GPS points. The velocity results and raw GPS and magnetometer measurements were used in a state fusion. Finally the iterative parameter estimation algorithm with varying covariances was executed on the raw measurements, the landmark based results and on the state fusion too. Because of the lack of reference wheel circumference the results could not be reliably compared. Therefore the raw measurements were analyzed deeply and approximate true values were calculated, in the case of straight movement. With this the result of the 3 cases was compared. These shows that the odometry fits very well to the landmark based path. The average position error is below than 1.5 m, which is a great result I think considering that the recursive odometry has been used for over 800 meters in a real case. The deviation from the calculated reference is only 7 mm, which is about 0.5%, but it decrease to 1 mm if the sideslip is used too. Therefore the effect of the sideslip in this type of estimation is also significant.