Translation-based KLT tracker under severe camera rotation using GPS/INS data

The Kanade-Lucas-Tomasi (KLT) faces a significant challenge with a translation model when the camera undergoes severe rotation. Although the use of an affine model can overcome this challenge, it is computationally expensive. In this letter, two solutions are proposed. Given reliable GPS/INS data, we determine the rotation angle between a stereo frame and use this knowledge to enhance the translation-based KLT tracker to remain robust under this situation. The experimental results on five pairs of UAV images show that the proposed methods can achieve equivalent tracking efficacy as the affine model while the computational cost remains close to that of the translation model.