Heteroscedastic noise estimation in Kalman filtering applied to road geometry estimation

Road geometry estimation is essential for self-driving vehicles and modern advanced driver-assistance systems (ADAS). State-of-the-art techniques utilize a Kalman filter to perform road geometry estimation. A common assumption in these Kalman filters is that the process- and measurement noise covariances are constant over time. However both sensor performance and process dynamics may change over time in real-world applications. Sensor performance may be affected by environmental factors such as rain and lighting conditions and process dynamics may depend on the road type. Noise processes like these with a feature dependent covariance are known as heteroscedastic noise. By estimating the heteroscedastic process- and measurement noise covariances more accurately both the filter performance and state uncertainty estimation may improve. Road geometry estimation is an especially interesting application in which to apply heteroscedastic noise estimation as there are several factors which intuitively should affect the process- and measurement noise. In this thesis a framework for heteroscedastic noise estimation in Kalman filtering applied to road geometry estimation is presented. The framework consists of two parts; a feature selection part and a heteroscedastic noise model. This noise model is constructed offline based on data set of ground truth state vector data. Two different state-of-the-art approaches for heteroscedastic noise modeling, a parametric approach and an approach based on a deep neural network, are evaluated as to determine if they are suitable for the application of road geometry estimation. Furthermore a straightforward approach that models heteroscedastic noise by dividing the features into discrete cases is studied. The noise models are quantitatively evaluated using a likelihood measure and root mean square error of the road geometry estimation. The results show that heteroscedastic noise estimation may improve both filter performance and estimation uncertainty consistency in the application of road geometry estimation.