Deep Features in Correlation Filters for Thermal Image Tracking

Object tracking using thermal infrared cameras has specific properties and challenges which distinguish it from the commonly used visual tracking. Recently, correlation filters (CF) based on deep features have been successfully applied in certain visual tracking scenarios. In this paper, we demonstrate that the success of these methods essentially depends on the way of how the deep features have been obtained. Indeed, the trackers based on CF and deep features use the pre-trained networks, originally trained for the object classification problem; hence, the obtained features are not invariant to changes of object appearance which may result from the change of camera type. We show that CF trackers based on deep features obtained from a convolutional architecture, pre-trained for visual object classification problem, have relatively poor performance when applied to the thermal tracking problem. Specifically, we test the performance of Kernelized Correlation Filter (KCF) on several chosen thermal video datasets, and demonstrate that the tracking results, when using simple feature representations (HOG features), are better than when using the pre-trained deep features. The results suggest that improved architectures and training methods for deep features should be developed in order to get more robust CF trackers.