Efficient point feature tracking based on self-aware distance transform

Abstract

In this paper, we propose a Self-aware Distance Transform (SDT) for efficient templatebased point feature tracking. The proposed SDT encapsulates the relationship between autocorrelation coefficients and the distance from the best match; therefore, it can be used to automatically determine the size of a search region in each point feature. The proposed SDT returns the expected distance between the predicted position and the best match from a statistical viewpoint, which guarantees a certain level of successful tracking depending on the cross-correlation at the predicted position. If the SDT returns a large expected distance due to the abrupt motion of a feature or inaccurate prediction, we progressively expand the search region on a hexagonal lattice while also using the SDT to reduce unnecessary computations. The performance of the proposed tracking method based on the SDT was verified experimentally in terms of its accuracy, robustness, and computational efficiency by comparing the proposed method to other tracking methods.