公共安全治理中可疑目标变形Transformer跟踪器

Suspicious Target Transformer Tracker in Public Security Governance

近年来,面向公共场所中智能视频监控的可疑目标跟踪算法取得一定的研究进展,有利于公共安全治理与应急事件的防范。凭借Transformer模型具有完整的Attention机制,本文尝试将其应用于公共安全治理中可疑目标持续精确跟踪与定位。具体为:分析可疑目标跟踪任务的特点,探索目标模板框架和搜索框架之间新颖的交互,提出面向公共安全治理中可疑目标可变形Transformer跟踪算法,命名为DeTrack,其分别构建基于可变形注意机制的编码器模块和基于自注意机制的编码器模块,并利用它们的组合来进行特征交互,之后利用所构建的角点预测头定位可疑目标。DeTrack无需关注所有像素即可定位目标,减少了模型参数量,在LaSOT、TrackingNet、GOT-10K和VOT2020上取得了较好的跟踪性能。

In recent years,certain research progress has been made on suspicious target tracking algorithms for intelligent video surveillance in public places,which is beneficial to public security governance and emergency incident prevention.With the complete attention mechanism in Transformer,we attempt to apply it to the continuous and accurate tracking and positioning of suspicious targets in public security governance.Specifically:Analyze the characteristics of the suspicious target tracking task,explore the novel interaction between the target template framework and the search framework,and propose a deformable transformer tracking algorithm for suspicious targets in public security governance,named DeTrack,which is based on the deformable attention mechanism.The encoder module and the encoder module based on the self-attention mechanism are used,and their combination is used for feature interaction,and then the developed corner prediction head is used to locate the suspicious target.DeTrack can locate targets without paying attention to all pixels,reduces the amount of model parameters,and achieves better tracking performance on LaSOT,TrackingNet,GOT-10K and VOT2020.