摘要
针对电子监控视频中受遮挡目标识别难度高的问题,提出一种电子监控部分遮挡目标单模态自监督信息挖掘技术。为了得到目标的状态信息,利用遮挡检测方法判断监控视频中是否存在部分遮挡目标。当监控视频存在部分遮挡目标时,利用减法聚类方法进行特定目标的识别、跟踪或描述,并提供更准确和详细的目标特征信息。在此基础上,将交叉熵损失函数与软间隔三元组损失函数构建的监督遮挡目标特征学习判别损失函数作为部分遮挡目标信息挖掘的目标函数,在每个批次的电子监控样本中,搜寻最小距离的负样本对以及最大距离的正样本对,并通过反向传播优化参数。由此输入电子监控图像样本,通过前向传播输出得到电子监控部分遮挡目标单模态自监督信息挖掘结果。实验结果表明,所提出的技术可以有效挖掘电子监控部分遮挡目标,目标挖掘的mAP值高于0.9,能够为提升监控目标识别精度提供可靠依据。
In allusion to the difficulty of identifying occlusive targets in electronic surveillance video,the single-mode self-supervised information mining technology for partially occlusive targets in electronic monitoring is researched.In order to obtain the state information of the target,the occlusion detection method is used to judge whether there is a partially occluded target in the surveillance video.When there are partially occluded targets in surveillance video,subtraction clustering method is used to identify,track or describe specific targets,and provide more accurate and detailed target feature information.On this basis,the supervised occlusion target feature learning discrimination loss function constructed by cross entropy loss function and soft interval triplet loss function is used as the objective function for partially occluded target information mining.In each batch of electronic monitoring samples,the minimum distance negative sample pair and the maximum distance positive sample pair are mined,and the parameters are optimized by the backpropagation.From this,the sample of electronic monitoring image is input,and the results of self-supervised information mining are obtained by means of the forward propagation.The experimental results show that this technology can effectively mine the partially occluded targets of electronic monitoring,and the mAP value of target mining is higher than 0.9,which can provide a reliable basis for improving the recognition accuracy of monitoring targets.
作者
周艳秋
高宏伟
何婷
辛春花
ZHOU Yanqiu;GAO Hongwei;HE Ting;XIN Chunhua(Inner Mongolia Agricultural University,Baotou 014109,China)
出处
《现代电子技术》
北大核心
2024年第10期47-51,共5页
Modern Electronics Technique
基金
内蒙古自治区科技计划项目(2020GG0033)
内蒙古农业大学职业技术学院项目(TDS202311)。
关键词
电子监控
遮挡检测
单模态自监督
信息挖掘
交叉熵损失函数
三元组损失函数
electronic monitoring
occlusion detection
single-mode self-supervision
information mining
cross entropy loss function
triplet loss funtion
