EM-based Extended Target Tracking with Automotive Radar using Learned Spatial Distribution Models

This paper presents a novel interpretation of data driven extended target tracking with applications to the automotive sector. Specifically, learning the spatial distribution of measurements from a vehicle in the form of a Variational Gaussian Mixture (VGM) model is examined. This distribution yields an interpretation applicable for the Expectation Maximization (EM) algorithm such that a closed-form measurement update for tracking an extended target can be derived. The approach is in particular designed for sparse and noisy measurements and is applied to Radio Detection and Ranging (RADAR) point information. Furthermore, an evaluation based on data from the recent nuScenes data set is performed.