End-To-End Decoupled Multi-Object Tracking of UAV Images Based On Scale Sensing

Multi-object tracking (MOT) is a key problem in drone image processing. Existing methods typically detect targets in each frame and then match the same target across frames through data association. However, these methods face several challenges in terms of performance improvement. On one hand, targets in drone images exhibit large scale variations, and traditional fixed-scale anchor boxes cannot adapt to these changes, leading to suboptimal detection performance. On the other hand, mainstream end-to-end multi-object tracking methods rely on large amounts of training data, but drone image datasets are often limited, which constrains tracking performance. To address these issues, this paper first proposes a scale-aware dynamic anchor box detection method, which significantly improves detection performance by estimating target scales and adjusting anchor boxes through a small convolutional neural network. Building on this, we further propose an end-to-end multi-object tracking method with decoupled detection and tracking, and apply a pretraining-finetuning strategy to mitigate the problem of limited data. Experimental results demonstrate that the proposed method significantly improves multi-object tracking performance on the VisDrone and UAVDT datasets, validating its effectiveness and robustness.