Detecting tracking errors via forecasting

We propose a tracker-independent framework to determine time instants when a video tracker fails. The framework is divided into two steps. First, we determine tracking quality by comparing the distributions of the tracker state and a region around the state. We generate the distributions using Distribution Fields and compute a tracking quality score by comparing the distributions using the L1 distance. Then, we model this score as a time series and employ the Auto Regressive Moving Average method to forecast future values of the quality score. A difference between the original and forecast returns an error signal that we use to detect a tracker failure. We validate the proposed approach over different datasets and demonstrate its flexibility with tracking results and sequences from the Visual Object Tracking (VOT) challenge.