Spatio-temporal Gaussian process models for extended and group object tracking with irregular shapes

Extended object tracking has become an integral part of many autonomous systems during the last two decades. For the first time, this paper presents a generic spatio-temporal Gaussian process (STGP) for tracking an irregular and non-rigid extended object. The complex shape is represented by key points and their parameters are estimated both in space and time. This is achieved by a factorization of the power spectral density function of the STGP covariance function. A new form of the temporal covariance kernel is derived with the theoretical expression of the filter likelihood function. Solutions to both the filtering and the smoothing problems are presented. A thorough evaluation of the performance in a simulated environment shows that the proposed STGP approach outperforms the state-of-the-art approach, with up to 90% improvement in the accuracy in position, 95% in velocity and 7% in the shape, while tracking a simulated asymmetric non-rigid object. The tracking performance improvement for a non-rigid irregular real object is up to 43% in position, 68% in velocity, 10% in the recall and 115% in the precision measures.