TrafficEKF: a Learning Based Traffic Aware Extended Kalman Filter

Most vehicle tracking algorithms only consider the vehicle’s kinematic state but ignore the information about the surrounding environment, which also plays an important role affecting how the driver controls the vehicle. In addition, how to represent the traffic information and its effect on the vesicle’s state is a challenging problem. In this paper, we propose a tracking method called traffic aware extended Kalman filter (TrafficEKF), which not only incorporates the vehicle’s kinematic dynamics, but also the information from the surrounding environment. The traffic information has been represented by a bird-eye-view rasterized image, with the road shape, traffic light conditions, and other objects inside the field of view. The effect of the traffic information on vehicle driving is learned by TrafficEKF from the ground truth data. Through training, the algorithm learns to predict the control input to the vehicle and to optimize the process and measurement noise covariance matrices used by the EKF. Based on experiments with real data, we show that the TrafficEKF significantly outperforms both a manually tuned EKF, and a data trained EKF, which ignore the environment information.