The improvement of KLT for real-time feature tracking from UAV image sequence

In the recent few decades, the demand on surveillance applications has been increasing significantly. Many of them involve identifying and tracking objects of interest in sequence of images. The Kanade-Lucas-Tomasi (KLT) is one of the known tracking techniques that has gained much interest in the field of motion tracking. We began our study by performing a preliminary experiment to measure the performance of KLT. The results suggested that the tracker can identify corresponding features in consecutive images greater than 80% accuracy. However, under large displacements or significant difference in illumination situations, KLT was unable to perform its task effectively. Our goal is to improve the tracking accuracy of KLT using a technique based on exterior orientation. By supplying the initial guessed positions of the corresponding points that is sufficiently close to the true values, the pyramidal KLT tracking process is expected to correctly converge to the final solutions at a far less depth levels. In addition, we propose a simple but helpful technique for radiometric adjustment based on exterior orientation.