Nighttime Air Tracking Based on Improved Unsupervised Siamese Network
Haoran Wei, Yanyun Fu, Deyong Wang, Rui Guo, Xueyi Zhao, Jian Fang · 2024
The limited light at night reduces the contrast between the target and its surroundings, making it difficult for the tracking algorithms to distinguish the target from similar looking objects or background elements. As the UAV rotates, the speed of movement causes the appearance characteristics of the target to change slightly, and it is very difficult to label the dark night dataset at this stage. Therefore, we design an end-to-end tracking algorithm, TransffCAR, which first pre-processes the video frame images to increase their brightness and locate potential targets from a large number of unlabelled video frames to generate training patches, avoiding the need for tedious manual labelling. The feature-extracted video frames are fed into the Transformer connection layer, which is used to assist in capturing global contextual information from aerial imagery, and then into the Dynamic Template Tracking module for updating to adapt to changes in the target's appearance, followed by adaptive hierarchical feature fusion, which improves the ability to perceive the target's positional information, and the target's feature information is corrected to facilitate subsequent input into the area regression network. Generate clear predictive tracking maps. We use NAT2021, NAT2021L and UAVDark70 to test on public datasets and test three evaluation metrics that outperform other state-of-the-art tracking methods. The experimental results confirm the robustness of our designed method in handling the task of nighttime UAV aerial video tracking.