An Adaptive Kernel based Correlation Filter Algorithm for Real Time Object Tracking

We proposed an adaptive kernel based correlation filter algorithm (AKCF) for object tracking. The algorithm includes two filters: a translation filter and a scale filter. In our method, we classify objects into three categories: fast speed ones, middle speed ones, and low speed ones. To track an object with different speeds, a “padding” value varies accordingly for adjusting the searching area in the translation filter. After detecting an object, the two filters update their parameters for dealing with the appearance variations and scale changes. In the strategy, the response score in the second frame is remembered as a reference, and an updating range is determined. When the current response score is out of the updating range, the filters stop updating to avoid introducing more interference from background. Finally, the proposed method is evaluated on the OTB dataset. Contrary to other state-of-the-art approaches, our method has provided good performance to track an object with different speed and serious scale changes. Additionally, our method is computational efficient.