Object tracking with convolutional neural networks and kernelized correlation filters

Convolutional neural networks are widely used in object recognition and detection. In recent years, some researchers attempt to apply deep neural networks to visual object tracking. However, deep networks are extremely time-consuming and object tracking is not a classification problem essentially. In this paper, we present an online tracking framework which combines shallow convolutional neural networks with kernelized correlation filters(KCF). Different from offline training, our method successfully gets the convolution kernels by K-means clustering algorithm. Experimental results based on a representative visual tracker benchmark dataset show that the proposed method achieves excellent performance.