Track-before-detect on radar image observation with an adaptive auxiliary particle filter
Audrey Cuillery, François Le Gland · 2019
This paper presents a multitarget tracking solution for 2D multitarget tracking from image observations including its application on radar data and results on simulated data. Tracking from image observations rather than from detected points is often referred to as track-before-detect (TBD). The objective is to capture targets with a low signal-to-noise ratio (SNR) which would not be detected or tracked after data thresholding. The highly nonlinear filtering equations are approximated using a particle filter implementation. This nonlinearity emerges from the observation function which relies on the radar treatment chain, involving matched filtering of a chirp signal, and beamforming achieved from a linear phased array, leading to the raw data image. Amplitudes and phases of target-returned signals are considered as temporally fluctuating, random and unknown, creating non-deterministic contributions of the targets to the signal to deal with. The adaptive auxiliary particle filter is presented, and is an extension of the auxiliary particle filter to the multitarget case under some independence assumptions. Auxiliary particle filter draws strength from its ability to imitate the optimal importance density for sampling new particles with a measurement based strategy, by resampling the promising particles. While targets are dynamically assumed independent, their individual region of influence on the radar image can overlap, which creates dependencies for the likelihood computation. Treating the whole radar image containing many cells is computationally very demanding. To reduce this expense, an observation gating is performed. Eliminating measurements which are far from the predicted measurements is a point measurement reasoning, and selecting a determined number of cells in the neighborhood of the target seems to be a drastic solution. We select here the informational cells from the a priori knowledge of the signal to capture more information, including side lobes induced by the beamforming process. The filter results are compared to other track-before-detect algorithm based on Markov chain Monte Carlo (MCMC).