A Kernel Correlation Filter Tracker with Fast Scale Prediction

Most existing scale solutions of Correlation Filter-based trackers fail to consider the priority of target scale calculation. Their high complexity destroys the high speed performance of the tracking method. To tackle this problem, an optimization strategy of combining scale prediction and scale pool is proposed. This method predicts the direction of future scale change by collecting the information of target historical scale variations, further divides the scale calculation priority, and uses the response characteristics of the correlation filter tracker to determine optimal scale. Both quantitative and qualitative experiments show that the proposed approach can achieve fast estimation of the target scale. The high real-time performance provides a guarantee for the migration of correlation filter target tracking algorithm to the development board with low computing capacity of embedded system.