Estimating Uncertainties of Recurrent Neural Networks in Application to Multitarget Tracking

Abstract

In multitarget tracking, finding an association between the new measurements and the known targets is a crucial challenge. By considering both the uncertainties of all the predictions and measurements, the most likely association can be determined. While Kalman filters inherently provide the predicted uncertainties, they require a predefined model. In contrast, neural networks offer data-driven possibilities, but provide only deterministic predictions. We therefore compare two common approaches for uncertainty estimation in neural networks applied to LSTMs using our multitarget tracking benchmark for optical belt sorting. As a result, we show that the estimation of measurement uncertainties improves the tracking results of LSTMs, posing them as a viable alternative to manual motion modeling.