Adaptive Kernel background intensity estimation based on local 2D orientation

Many target tracking algorithms for radar systems assume homogeneous backgrounds of clutter. However, real backgrounds are rarely homogeneous. By estimating background intensity, and using the estimate in the likelihood measure, the tracking algorithm is given the ability to adapt to the background. In this work, a method for estimating the clutter intensity is introduced. The method is based on locally adaptive Kernel Density Estimation (KDE), where local 2D structure of the background in terms of energy and orientation controls the smoothing properties of the filter kernels. In regions with low clutter intensity, the kernel adopts low-pass characteristics, and the intensity estimate is based on observations from a larger volume. In regions where there are ridges in the clutter intensity, kernels are selected such that smoothing is carried out along ridges instead of across them. Peaks in the clutter intensity are left unsmoothed. The proposed method is compared to other methods on synthetic data. Additionally, a demonstration is given on recorded radar data.