Cooperative Plug-and-Play-KalmanNet for 4D situational awareness in autonomous driving

Fully operational autonomous driving requires accurate 4D situational awareness from vehicles, i.e., 3D location information over time. For that purpose, each vehicle establishes connectivity links with its direct neighboring vehicles, collects, exchanges and fuses multi-modal measurements in order to increase the accuracy of localization information. In this work, we focus on efficiently tracking the unknown system dynamics, and as a matter of fact boosting the performance of 4D situational awareness, without knowing beforehand the statistics of sensor measurement noise. More specifically, we exploit the so-called KalmanNet, a recurrent neural network, which has been explicitly designed to estimate the covariance and Kalman gain matrices of standard Extended Kalman Filter through a typical deep learning approach. Pre-trained KalmanNet is then used as the core module of a cooperative localization approach which effectively realizes dynamic 4D situational awareness and improves sensor fusion performance. An extensive evaluation study in CARLA autonomous driving simulator demonstrates that our cooperative plug-and-play approach significantly improves localization accuracy, as well as reduces drift error which is a critical error metric in the automotive industry.