Hide/Show Apps

Non-parametric sensor bias estimation using gaussian process

Deniz, Ufuk Mehmet
In this work, we propose to use Gaussian Process (GP) regression techniques to estimate possible non-parametric sensor biases in a multi-sensor environment. Using multi-sensor for tracking applications makes a system more reliable. Combining information acquired from multiple sensors has several difficulties. In a multi-sensor environment, it is difficult to assure that all sensors are registered and calibrated perfectly during operation. When the error between different sensors exceeds tolerable limits, multi-sensor tracking might become a problem. In this study, we assume that the biases between local agents stem from unidentified and complicated sources, which would make parametric modeling very difficult, if not impossible. GPs are used to model the unknown bias that may occur between the measurements/estimates of two different sensors. The model does not require a parametric model for the bias error structure. Since standard GP becomes impractical in reality because of the computational burden of large data sets, a sparse approximation of GP is implemented to scale down the computational complexity. Estimated biases can later be used to perform association or fuse tracks among sensors more accurately. We present a review of three different data fusion architectures so as to combine the data from multiple sensors. We compare the tracking performance of the data fusion architectures to explore the one which provides the best accuracy in our study.