A Scale Adaptive Kernel Correlation Filter-Based Tracker with Optimized Update Strategy

Correlation filter-based trackers have achieved competitive results in the field of visual tracking, but the abilities to deal with the multi-scale and keep the tracking stability still need to be improved. In this article, the complex background tracking problems are studied based on the kernel correlation filter. In order to address the fixed size limitation, we present an improved scheme, which can not only realize scale adaption, but also retain the advantages of the original algorithm. In addition, the adaptive update learning rate of the model is achieved by using the peak sidelobe ratio as an indicator to measure the filter performance. Using the index, we can selectively reduce the model interference to low precision information. Moreover, we design a redetection mechanism that runs automatically when the tracking evaluation indicator is low. Compared with the radical model update methods, those conservative update strategies can effectively decrease the model pollution and ensure the tracking effect. Extensive verifications based on OTB-100 benchmark show that the optimized algorithm improves the accuracy, robustness, and generalization ability of target tracking.