State Estimation for Autonomous Vehicle

Abstract

The ability to comprehend the driving scenario is a prerequisite for autonomous driving. It is the preliminary stage for the motion planning system in the control routine of an autonomous vehicle. To offer estimations regarding the movements of the ego-vehicle and nearby obstacles, estimation algorithms must manage a large amount of data from various sensors. Furthermore, tracking is critical in estimating the condition of obstacles since it ensures that impediments are recognized over time. In this project we developed an estimator for Autonomous vehicle, using error state-extended kalman filter and analyzed its results in comparison with a standard linear kalman filter. Kalman Filters have played major role in the development of many technologies like space crafts and self-driving cars, they are used to estimate their trajectories. It’s a need to get rid of the false data or noise in sensors measurements, in order to enhance the safety and lower the risk of accidents. So that depicts the importance of state estimation. Performed estimation on three local quantities, position of car in ENU frame, linear velocity with which car was moving and orientation of car. For which we collected data of three sensors (IMU, GNSS and LIDAR), using CARLA, in simulated environment, containing vehicles and pedestrians. Kalman filters are basically based on the phenomenon of sensor fusion, collecting data from each sensor and then fusing the data to fetch the best estimate for a particular state. ES-EKF Setup is based on two models, firstly motion model which used IMU data to predict the local states of vehicle and then a measurement model, which used GNSS and Lidar data to correct the predictions. Performed five types of validation tests on our algorithm for both linear kalman filter and non-linear kalman filter (ES-EKF) and analyzed their results with respect to root mean square error (RMSE). In all the cases, the non-linear kalman filter gave less RMSE as compared to linear kalman filter, which shows that non-linear system estimates the states and localizes the Autonomous vehicle accurately.