Research on an Improved KCF Target Tracking Algorithm Based on CNN Feature Extraction

Abstract

Target tracking is one of the most concerned computer problems, but it is also challenging with few training samples, fast moving objects and some other issues. The kernelized correlation filter (KCF) algorithm proposed by the team of Joao F. Henriques had applied to address this problem for tracking successfully. The method has expanded the number of negative samples to enhance the performance of the tracker and used the fast Fourier transform to accelerate the calculation of the algorithm. However, the features used by the KCF have limited ability to express the target with complex background. We propose improved KCF algorithm for tracking. The pre-trained deep convolutional neural network (CNN) is introduced in extracting the layer information respectively to describe the spatial and semantic features of the target. Experiments are performed on OTB-2015 benchmark datasets, and the results show that in comparison with the existing tracking algorithms, the proposed improved algorithm can deal with the challenges much better performance compared to original KCF and KCF-S method.