Particle Filter-Based Map Matching for Vehicle Localization Using Inertial and Odometry Sensors

Sammanfattning

This thesis investigates GNSS-denied vehicle localization using particle filter-based map matching with wheel odometry and inertial sensors. The main contribution is an evaluation of design choices affecting both localization accuracy and real-time performance, such as likelihood modeling, wheel radius estimation, and sensor choice. To do so, a standard particle filter was compared to variants with a single modification in four separate studies. The results show that particle filter-based map matching significantly improves localization accuracy compared to dead reckoning by constraining motion to road geometry. A laterally shifted Gaussian likelihood model used in the measurement update improves accuracy compared to a centered Gaussian assumption by better reflecting right-lane driving behavior. Furthermore, augmenting the state vector with wheel radius estimation reduces errors in the motion model and improves long-term accuracy. From a computational perspective, reduced update rates improve efficiency without degrading performance within a stable operating range. The Auxiliary particle filter does not provide consistent benefits over the standard particle filter in this setting, particularly at higher particle counts. Regarding sensor choice, wheel encoder and wheel-mounted IMU configurations achieve comparable accuracy, while chassis-mounted IMU setups perform worse due to double-integration drift. Overall, the study demonstrates that accurate real-time GNSS-free localization is feasible using particle filter-based map matching, where performance is primarily driven by motion and measurement modeling rather than increased filter complexity.