面向智能视频监控的人体小目标检测

Tiny Person Detection for Intelligent Video Surveillance

人体目标检测对社会治理和城市安全具有很重要的现实意义,监控数据是数据安全的重要来源。小目标检测是目前受到广泛关注的安全检测问题中一项具有挑战性的任务,其检测对象为大型图像中少于20个像素的目标。小目标的特征难以表征,其中一个主要挑战是,用于预训练/共同训练检测器的数据集(如COCO)与用于微调检测器的数据集(如TinyPerson)之间存在尺度不匹配的情况,这给小目标检测器的性能带来了负面影响。为了解决这个问题,文中提出了一种优化策略,用于匹配不同数据集的尺度,称其为尺度分布搜索(Scale Distribution Search,SDS),同时平衡图片的信息收益(数据集之间的尺度相近)和信息损失(信噪比(SNR)的降低)。该策略使用高斯模型对数据集中目标的尺度分布进行建模,通过迭代的方式寻找最优分布参数;并对比数据集中目标的特征分布和检测器的性能,以找到最佳的尺度分布。通过SDS策略,主流目标检测方法在TinyPerson上实现了更好的性能,证明了SDS策略在提升预训练/共同训练效率上的有效性。

Person detection has significant practical implications for social governance and urban security.Monitoring data is an important source of data security.Tiny object detection,which focuses on less than 20 pixels objects in large-scale images,is a challenging task.One of the main challenges is the scale mismatch between the dataset used for pre-training/co-training the detectors,such as COCO,and the dataset used for fine-tuning the detectors,such as TinyPerson,which negatively affects the performance of detectors on tiny object detection.To address this challenge,this paper proposes an optimization strategy called scale distribution searching(SDS)to match the scale of different datasets for tiny object detection,which also balance the information gain and loss.The Gauss model is used to model the scale distribution of targets in the dataset,and the optimal distribution parameters are found through iteration.The feature distribution and the performance of the detector is comparedto find the best scale distribution.Through the SDS strategy,mainstream object detection methods have achieved better performance on TinyPerson,demonstrating the effectiveness of the SDS strategy in improving pre-training/co-training efficiency.