Convolutional neural network with structural input for visual object tracking

Numerous deep learning approaches have been applied to visual object tracking owing to their capabilities to leverage huge training data for performance improvement. Most of these approaches have limitations with regard to learning target specific information rich features and therefore observe reduced accuracy in the presence of different challenges such as occlusion, scale variations, rotation and clutter. We proposed a deep neural network that takes input in the form of two stacked patches and regresses both the similarity and the dis-similarity scores in single evaluation. Image patches are concatenated depth-wise and fed to a six channel input of the network. The proposed network is generic and exploits the structural differences between the two input patches to obtain more accurate similarity and dissimilarity scores. Online learning is enforced via short-term and long-term updates to improve the tracking performance. Extensive experimental evaluations have been performed on OTB2015 and TempleColor128 benchmark datasets. Comparisons with state-of-the-art methods indicate that the proposed framework has achieved better tracking performance. The proposed tracking framework has obtained improved accuracy in different challenges including occlusion, background clutter, in-plane rotation and scale variations.