Comparison of Track to Track fusion methods for nonlinear process and measurement models

Abstract

Automotive sensors play a vital role in the environment perception for vehicles in advanced driver assistance systems (ADAS). Sensors have their own distinctive advantages and drawbacks, which makes it imperative to fuse information from disparate sources. The fusion can be performed either at the sensor or the track level. Track to track fusion (T2TF) offers a big advantage as individual sensor blocks can be treated as grey or even black boxes i.e. a very limited knowledge of their characteristics might be required. In this paper, we study T2TF for a single target vehicle, tracked by two generic sensors, differing in kinematic tracking accuracy. A challenging reference trajectory is simulated consisting of both linear and nonlinear motion segments. The main objective is to compare the performance of different nonlinear sensor fusion algorithms, comprising of several combinations of prediction and measurement update methods. We show that the covariance intersection based update methods outperform the Kalman filter derivatives, as they tend not to produce overly optimistic estimates.