High-performance UAVs visual tracking using deep convolutional feature

The application of visual tracking down unmanned aerial vehicles (UAVs) is an important research direction. Although many existing UAVs visual trackers exploit the features of deep convolution to effectively improve the robustness of trackers, the target features extracted by convolutional neural network (CNN) are difficult to distinguish when facing occlusion, illumination variation, viewpoint change, deformation, and scale variation. Especially for distractors (such as similar objects), these trackers cannot capture temporary appearance changes. In this work, we propose an efficient UAVs visual tracker, which can effectively alleviate the impact of occlusion, viewpoint change, and illumination. First, we stretch the width of the network to acquire affluent target appearance feature information. Then, we design an attention information fusion module (AIFM) to enhance feature extraction, which can effectively establish the correspondence relationship of long-range pixel pairs between the template frame and the detection frame. The ability of the tracker to distinguish the target can be effectively improved through suppressing the global background response. Furthermore, we design a multi-spectral information fusion module (MSIFM) to dynamically learn the appearance features of the detection frame target corresponding to the template frame features, which can improve the prediction accuracy of the bounding box. Finally, the distance intersection over union is employed to evaluate the object location and complete the prediction of the bounding box. Abundant experiments demonstrate that the proposed method has powerful tracking performance in a diversity of UAVs scenarios.