Infrared target tracking algorithm combining deep features and gradient features

Due to the low contrast between the target and the background of the infrared sequence image, the edge of the image is blurred and the dynamic range of the gray level is small, what features is used to describe the target becomes the key to tracking. Deep features and gradient features are the main features of most current tracking algorithms. However, the target semantic of deep feature extraction pays attention to intra-class classification, ignores intra-class differences, and is easily interfered by similar background (distractor); gradient features as a local area feature, it is not susceptible to background interference, but it cannot adapt to the dramatic deformation of the target. Based on the complementarity of these two features, this paper proposes an infrared target tracking algorithm that combines deep features and gradient features. In this paper, deep features and gradient features are used to represent the semantic and local structure of the target respectively, which enhances the ability to represent arbitrary targets. Next, the tracking model established by different features further improves the robustness of tracking. Finally, this paper establishes a model mutual aid mechanism, and uses the complementarity between the deep feature tracking model and the gradient feature tracking model to accurately target. In the experiment, this paper selects the latest infrared video tracking database (VOT-TIR2016) to verify the effectiveness of the proposed algorithm. The results show that compared with the current mainstream tracking algorithm, the algorithm achieves a 3.8% improvement in accuracy and a 4.3% improvement in success rate, it can effectively handle the effects of similar background and deformation in tracking.