Assessing the accuracy of image tracking algorithms on visible and thermal imagery using a deep restricted Boltzmann machine

Abstract

Image tracking algorithms are critical to many applications including image super-resolution and surveillance. However, there exists no method to independently verify the accuracy of the tracking algorithm without a supplied control or visual inspection. This paper proposes an image tracking framework that uses deep restricted Boltzmann machines trained without external databases to quantify the accuracy of image tracking algorithms without the use of ground truths. In this paper, the tracking algorithm is comprised of the combination of flux tensor segmentation with four image registration methods, including correlation, Horn-Schunck optical flow, Lucas-Kanade optical flow, and feature correspondence methods. The robustness of the deep restricted Boltzmann machine is assessed by comparing between results from training with trusted and not-trusted data. Evaluations show that the deep restricted Boltzmann machine is a valid mechanism to assess the accuracy of a tracking algorithm without the use of ground truths.