An adaptive kernelized correlation filters with multiple features in the tracking application

Abstract

Automatic target detection and tracking systems are used extensively in complex scenes. In long-term tracking, some visual attributes of objects are changing, such as illumination, size, profile, and so on. To address the issue, it is particularly important to describe the essential properties of the objects in tracking. An enhanced kernelized correlation filter tracking strategy fused multiple features with location prediction is proposed. To make the object appearance models more accuracy and robustness, based on the original histogram of oriented gradient features, we integrate the hue, saturation, value, and grayscale information to construct a new descriptor to represent the target appearance. Moreover, location prediction and bi-linear interpolation are employed to obtain the more accurate target position. Experiments show that the proposed strategy can obtain superior or competitive performance in challenging benchmark data sets. In practice, the algorithm is applied to track shuttle bus targets in the airport apron.