Aerial Small Object Tracking with Transformers
Chang Liu, Sheng Hua Xu, Baochang Zhang · 2021 IEEE International Conference on Unmanned Systems (ICUS) · 2021
Due to the long distance of UAV aerial photography and the small proportion of objects, small object tracking represented by UAV aerial photography has always been a challenging part in the tracking field. Through experiments, we found that such challenges are strongly correlated with attributes such as occlusion, out-of-view, and drift. In this paper, we applied trackers with transformer and template update to aerial small object tracking, and applied them to the UAV aerial photography datasets and the small object datasets to propose corresponding small object trackers. The attention module in the transformer can provide a global response between template frame and search frame, which can effectively supplement contextual information for small objects, and solve the problem that small objects have weak feature expression ability. Using the template update strategy, the update of the template can be controlled under the conditions of occlusion, out-of-view, and drift that small objects are prone to appear, so the performance of the trackers on the small objects can be improved. Experiments show that these two small object trackers can achieve state-of-the-art performance on the corresponding datasets, and the improvements in attributes occlusion, out-of-view, and drift are particularly obvious.