Abstract:

The collection of UAV (unmanned aerial vehicle) video datasets is costly and faces issues such as limited quantity, low quality, and scenario constraints. To address these challenges, an unsupervised UAV-object-tracking model based on temporal cycle consistency and dynamic memory enhancement was proposed. First, salient-object detection was introduced for unlabeled object discovery. By combining salient object detection with unsupervised optical flow techniques and incorporating dynamic programming based on image entropy, the quality of pseudo-labels was improved. Second, a weight is defined for each frame in the video, and these weights are utilized for single-frame training to fully leverage the information from all frames. Finally, inspired by long short-term memory (LSTM) networks, the memory queue was transformed into a dynamic memory queue, along with a self-attention branch designed to control its updates. Target-features changes over long spans were learned without increasing the queue length. The proposed method achieved 68% accuracy on UAV datasets, outperforming other unsupervised trackers and matching typical supervised-tracker performance. On general scene datasets, it attained 77?% accuracy, comparable to other unsupervised trackers. Experimental results on both UAV and general scene datasets demonstrated that the proposed method achieved excellent performance in scenarios involving rapid motion and large-scale variations.

Key words: object tracking, unmanned aerial vehicle, unsupervised learning, attention mechanism, twin network