Several hot topics of high practical relevance in the context of recursive estimation of states and parameters based on data from multiple sensors will be discussed.
First, new types of nonlinear Kalman filters based on optimal deterministic sampling approximations of continuous density functions are shown. These filters provide an adjustable tradeoff between estimation quality and computational complexity. They also allow the optimal estimation of periodic quantities, such as angles or orientations.
Second, the direct fusion of two estimates characterized by data points only is discussed. This is a practically relevant problem, as often no further knowledge about the estimates is available. It arises, e.g., when fusing the output of two particle filters.
Third, when tracking multiple objects with high-resolution sensors, the problem of associating sensor data with objects and object parts arises. In this talk, I will argue for association-free methods that never explicitly associate sensor data with objects and provide high-quality estimates with a low computational complexity.
All methods are based on a new distance measure between probability density functions. This distance can handle a comparison between arbitrary continuous and discrete densities, in particular between empirical distributions. It can efficiently be computed and is continuously differentiable, which makes it useful for optimization purposes.
Some application examples such as object tracking for belt sorting of bulk material, person tracking in extended range telepresence systems, and surface reconstruction in beating heart surgery will be shown.