【摘要】 静态平台上的联合检测与嵌入范式多目标跟踪模型已取得显著进展，但在面对无人机(Unmanned aerial vehicle, UAV)遥感场景中快速运动、视角剧烈变化及高密度小目标聚集等挑战时，现有方法难以保持稳定的跟踪性能。为此，对不确定性感知无监督多目标跟踪(Uncertainty-aware unsupervised multi-object tracking, U2MOT)模型进行了改进，进一步提升其在无人机场景下的性能表现。在重识别预测头部分支中引入所提出的像素注意力聚合模块(Pixel attention aggregation module, PAAM),通过增强的自注意力机制以及稳定的双重残差结构，增强了目标身份特征的提取能力。针对无人机平台中目标运动变化剧烈的问题，在对比学习训练样本生成阶段构建了基于轨迹片段运动信息的数据增强策略，结合类别引导的透视变换模拟真实运动变化，增强了样本的多样性与判别性。在损失函数中引入多任务自适应加权损失机制，实现检测与重识别任务的动态权重调整，提高模型训练的稳定性与协同性。在视觉与无人机多目标追踪(Vision meets drone:multiple object tracking, VisDrone MOT)数据集上的实验结果表明，所设计的跟踪算法在多目标跟踪准确率和身份一致性得分上分别提升了2.1%和1.7%,有效增强了模型在无人机场景中的跟踪性能。更多还原

【Abstract】 Joint detection and embedding paradigms have achieved significant progress in multi-object tracking tasks on static platforms. However, existing methods often struggle to maintain stable tracking performance when confronted with challenges in drone remote sensing scenarios of unmanned aerial vehicle(UAV), including rapid camera motion, dramatic viewpoint changes, and densely distributed small targets. To address these challenges, this paper proposes an improved uncertainty-aware unsupervised multi-object tracking(U2MOT) model to further enhance its performance in unmanned aerial vehicle scenarios. A pixel attention aggregation module(PAAM) is introduced into the reidentification prediction branch to enhance identity feature extraction through an improved self-attention mechanism and a stabilized double-residual structure. To address the issue of highly dynamic target motion in drone platforms, a data augmentation strategy named Drone tracklet motion augmentation is proposed during the contrastive learning sample generation stage. This strategy leverages category-guided perspective transformations based on motion cues from short tracklets to enhance the diversity and discriminability of samples. An automatic weighted loss mechanism is employed in the loss function to dynamically balance the optimization of detection and reidentification tasks, thereby improving the stability and coordination of multi-task training. Experiments on the vision meets drone that multiple object tracking(VisDrone-MOT) dataset demonstrate that the proposed method improves the multi-object tracking accuracy and identity F1 score by 2.1% and 1.7%, respectively, effectively enhancing tracking performance in unmanned aerial vehicle scenarios.更多还原