Comparing EKF, UKF, and PF Performance for Autonomous Vehicle Multi-Sensor Fusion and Tracking in Highway Scenario

Abstract

Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), and Particle Filter (PF) are three popular algorithms to address obstacle position estimate and tracking problems. However, as technology develops, autonomous vehicles pursue a better understanding of the environment and higher safety driving. Different modern sensors are mounting on the car, such as three-dimensional Light Detection and Ranging (LiDAR) and Radio Detection and Ranging (Radar). Sensor fusion from various data types can improve the position estimate accuracy and challenge the traditional tracking algorithm. In order to explore which tracking algorithm has better performance in multi-sensor data fusion (MSDF) and multi-target tracking (MTT) problems, this paper implements and analysis EKF, UKF, and PF algorithm for an autonomous vehicle with three LiDAR and two RADAR in a highway scenario. Our first contribution is processing the point cloud data for each sensor and using a bounding box data type to normalize individual obstacles. Then we designed a tracking system that can switch EKF, UKF, and PF tracking algorithms. Third, we use different state vector update matrices for LiDAR and RADAR for position updates and speed updates. Actual highway driving data are recorded, and a Robotic Operating System (ROS) model is built for algorithm development and result analysis.