KCNN: A Kernelized Correlation Filter with Convolutional Neural Network Object Tracker

Object tracking is one of the classic problems in video and image processing and has numerous applications in computer-human interaction, security, traffic control, medicalimaging and other fields. Several challenges exist including the change in illumination variation, motion blur and fast motion of the target object. Hence, a tracking method should be able to not only track the object in sudden rotations and movements, but also when background clutters, with high precision. In this paper, we first use a pre-trained architecture of convolutional neural networks, named VGG-NET, to extract features from the layers with high-level semantic information. This helps discrimination ofobjects in the image. Then, the output of each layer is fed into a kernelized correlation filter to find the maximum response of each filter using a Coarse-to-Fine (CTF) search algorithm. Hence, the tracker approximates the location of the object in the current frame and this leads to higher precision and reliability values. Experimental results on a set of some challenging videos in the TB- 100 database show that the proposed method outperforms previous methods by reaching a precision up to 80%.

کلیدواژه ها:

Object Tracking ، Kernelized Correlation Filter ، Deep Neural Networks ، Convolutional Neural Networks