EVENT DOSSIER
To address the issue that existing methods perform well in fixed-camera scenarios but fail in mobile drone perspectives, researchers released the MovingDroneCrowd++ dataset and proposed two new methods: GD3A and GIA-Track. GD3A achieves video individual counting through global density map decomposition, while GIA-Track establishes multi-object tracking trajectories based on group identity associations. Both methods are integrated into a single framework, utilizing intra-group density distribution and aggregated pixel-level matching across frames, and optimizing optimal transmission through frame-pair conditional trash can scores to distinguish between different identities. Experiments show that this method reduces counting errors by 47.4% and improves tracking accuracy by 64.6% in mobile drone videos. The related code, dataset, and pre-trained models are now open-source.
研究人员发布 MovingDroneCrowd++ 数据集，旨在解决移动无人机拍摄下的密集人群计数与跟踪难题。针对现有方法在固定相机场景下表现良好但在无人机视角下失效的问题，团队提出了 GD3A 和 GIA-Track 两种方法。GD3A 通过全局密度图分解实现视频个体计数（VIC），GIA-Track 基于组内身份关联建立多目标跟踪（MOT）轨迹。两者统一于一个框架，利用组内密度分配和组内身份关联聚合跨帧像素级匹配，并通过帧对条件垃圾桶分数优化最优传输以区分身份差异。实验表明，该方法在移动无人机视频上使计数错误降低 47.4%，跟踪准确率提升 64.6%。相关代码、数据集及预训练模型已开源。