Vehicle Lateral and Longitudinal Velocity Estimation Using Machine Learning Algorithms

Prediction and estimation of states are of great importance for vehicle control and safety. The conventional observers are designed and used vastly in the automotive industry to fulfill this objective. However, the vehicle behavior can be nonlinear or unpredictable, and it is difficult or impossible to use linear or low-degree nonlinear observers to estimate vehicle states. These observers may fail to estimate the states correctly in high slippery roads, combined slip situations, and maneuvers with intense steering inputs. In this study, two kernel regression-based machine learning methods are used to estimate lateral and longitudinal velocities. In the estimations with kernel regression methods, only a limited number of reference points in the vicinity of estimation space are needed. The kernel-based methods do not need training and can be implemented for real-time applications. The estimation methods are capable of estimating lateral or longitudinal velocities with a frequency of more than 50 Hz. The suggested estimation methods can be applied for different vehicles without the need to be changed or modified. The proposed approach is capable of utilizing data of any vehicle by normalization. The resulting solution can be used for the state estimation of any vehicle. Since the estimation methods rely on the local reference points, the lack of rich reference points may be a challenge for these estimation methods. Two healing algorithms are proposed to address this issue and make the reference points richer in the vicinity of the estimation space. The performance of the healing algorithms for different maneuvers is also studied in this thesis. A series of simulation and experimental tests with various road conditions are utilized for validating the estimation performance. Results show that the proposed algorithm can estimate different vehicles' lateral and longitudinal velocities. The estimation methods can estimate the lateral velocity with an error of less than 0.1 m/s. The methods can estimate the longitudinal velocity with an error of less than 2 kph if the proper reference data is provided