Robust Infrared Air Object Tracking Fusing Convolutional and Hand-Crafted Features

The infrared objects do not have color information, and they have low resolution. Therefore, the hand-crafted features cannot robustly describe observation model of the object, and it is easy to track failure in the presence of heavy occlusion and infrared distractors. Based on the correlation filtering theory, a robust air object tracking algorithm using convolutional and hand-crafted features is proposed in this paper. Firstly, there are differences in the ability of different layer features to describe the objects. We reconstruct the foreground mask with feature map selection approach, and select the features which are sensitive to intra-class appearance variation. Then, convolutional and hand-crafted features are fused and embedded in the correlation filtering theory to estimate the object position, achieving the air object tracking. Finally, to re-capture the object when the tracking fails, the proposed algorithm introduces YOLOv3 for re-detection. We verify our algorithm with actual infrared image sequence and the simulation image sequence. The experimental results show that the proposed algorithm can accurately track air objects with high precision.