Comparison of Object Tracking Methods and Performance Analysis of Kernelized Correlation Filter with Different Appearance Models

Abstract

Object tracking enables determination of relevant objects in the moving images without the need for object detection in each frame. There are various studies that are based on correlation, deep learning and different methods in this field in the literature. In these studies, different appearance models such as Scale Invariant Feature Transform (SIFT), Histogram of Oriented Gradient (HOG), motion model, convolutional neural network features are used to model the object to be tracked. Having operations performed in Fourier space, the correlation approach is effective in making object tracking systems run in real time. Using the correlation filter with the convolutional neural network, the success of object tracking can be increased. The aim of this study is to compare different object tracking methods and analyze the effect of different appearance models on object tracking. Object tracking comparison (OTB-100) data set which includes 100 images series prepared in different environments and conditions, was used. Two measures of precision and success were computed for performance measurement. When the obtained results of the experiments on the OTB data set were evaluated, it was observed that the combined use of the correlation filter and convolutional neural network increased the object tracking performance.