Shallow Layer Convolutional Features with Correlation Filters for UAV Object Tracking

In this paper, convolutional shallow features are proposed for unmanned aerial vehicle (UAV) tracking. These convolutional shallow features are generated by pre-trained convolutional neural networks (CNN) and are used to represent the target objects. Furthermore, to estimate the location of the target objects, an adaptive correlation filter based on the Fourier transform is used. This filter is multiplied with the convolutional shallow features by using pixel-wise multiplication in the Fourier domain. Then, the inverse of Fourier is performed to estimate the location of the target object, where its location is represented by the maximum value of the response map. Unfortunately, the target object always changes its appearance during tracking. Therefore, we proposed an updated model to address this issue. The proposed method is evaluated by using the UAV123 10fps benchmark dataset. Based on the comprehensive experimental results, the proposed method performs favorably against state-of-the-art tracking algorithms