Improving the Targets’ Trajectories Estimated by an Automotive RADAR Sensor Using Polynomial Fitting

Abstract

A way to reduce the uncertainty at the output of a Kalman filter embedded into a tracker connected to an automotive RADAR sensor consists of the adaptive selection of parameters during the tracking process. Different informed strategies for automatically tuning the tracker’s parameters and to jointly learn the parameters and state/output sequence using: expectation maximization; optimization approaches, including the simplex algorithm; coordinate descent; genetic algorithms; nonlinear programming using finite differencing to estimate the gradient; Bayesian optimization and reinforcement learning; automatically tuning hyper-parameters in the least squares, were already proposed. We develop here a different semi-blind post-processing approach, which is faster and more robust. Starting from the conjecture that the trajectory is polynomial in Cartesian coordinates, our method supposes to fit the data obtained at the output of the tracker to a polynomial. We highlight, by simulations, the improvement of the estimated trajectory’s accuracy using the polynomial fitting for single and multiple targets. We propose a new polynomial fitting method based on wavelets in two steps: denoising and polynomial part extraction, which compares favorably with the classical polynomial fitting method. The effect of the proposed post-processing methods is visible, the accuracy of targets’ trajectories estimations being hardly increased.