目的卷积神经网络广泛应用于目标检测中,视频目标检测的任务是在序列图像中对运动目标进行分类和定位。现有的大部分视频目标检测方法在静态图像目标检测器的基础上,利用视频特有的时间相关性来解决运动目标遮挡、模糊等现象导致的漏检...目的卷积神经网络广泛应用于目标检测中,视频目标检测的任务是在序列图像中对运动目标进行分类和定位。现有的大部分视频目标检测方法在静态图像目标检测器的基础上,利用视频特有的时间相关性来解决运动目标遮挡、模糊等现象导致的漏检和误检问题。方法本文提出一种双光流网络指导的视频目标检测模型,在两阶段目标检测的框架下,对于不同间距的近邻帧,利用两种不同的光流网络估计光流场进行多帧图像特征融合,对于与当前帧间距较小的近邻帧,利用小位移运动估计的光流网络估计光流场,对于间距较大的近邻帧,利用大位移运动估计的光流网络估计光流场,并在光流的指导下融合多个近邻帧的特征来补偿当前帧的特征。结果实验结果表明,本文模型的m AP(mean average precision)为76.4%,相比于TCN(temporal convolutional networks)模型、TPN+LSTM(tubelet proposal network and long short term memory network)模型、D(&T loss)模型和FGFA(flow-guided feature aggregation)模型分别提高了28.9%、8.0%、0.6%和0.2%。结论本文模型利用视频特有的时间相关性,通过双光流网络能够准确地从近邻帧补偿当前帧的特征,提高了视频目标检测的准确率,较好地解决了视频目标检测中目标漏检和误检的问题。展开更多
General multi-protocol label switching(GMPLS) based on traffic engineering is one of the possible methods to implement all-optical network. This method implements the network with IP technique and guarantees the quali...General multi-protocol label switching(GMPLS) based on traffic engineering is one of the possible methods to implement all-optical network. This method implements the network with IP technique and guarantees the quality of service with traffic engineering. Based on the establishment of selecting schemes of optical path and methods of traffic calculation, the wavelength routing algorithm of all-optical network based on traffic engineering is presented by combining with prior route of shortest path and traffic engineering, the algorithm procedures are given, and the actual examples are introduced as well as the analysis on simulation calculation. This research results have certain significance for the achievement of optical switching technique of all-optical network.展开更多
目的 视频异常行为检测是当前智能监控技术的研究热点之一,在社会安防领域具有重要应用。如何通过有效地对视频空间维度信息和时间维度信息建模来提高异常检测的精度仍是目前研究的难点。由于结构优势,生成对抗网络目前广泛应用于视频...目的 视频异常行为检测是当前智能监控技术的研究热点之一,在社会安防领域具有重要应用。如何通过有效地对视频空间维度信息和时间维度信息建模来提高异常检测的精度仍是目前研究的难点。由于结构优势,生成对抗网络目前广泛应用于视频异常检测任务。针对传统生成对抗网络时空特征利用率低和检测效果差等问题,本文提出一种融合门控自注意力机制的生成对抗网络进行视频异常行为检测。方法 在生成对抗网络的生成网络U-net部分引入门控自注意力机制,逐层对采样过程中的特征图进行权重分配,融合U-net网络和门控自注意力机制的性能优势,抑制输入视频帧中与异常检测任务不相关背景区域的特征表达,突出任务中不同目标对象的相关特征表达,更有效地针对时空维度信息进行建模。采用LiteFlownet网络对视频流中的运动信息进行提取,以保证视频序列之间的连续性。同时,加入强度损失函数、梯度损失函数和运动损失函数加强模型检测的稳定性,以实现对视频异常行为的检测。结果 在CUHK(Chinese University of Hong Kong) Avenue、UCSD(University of California, San Diego) Ped1和UCSD Ped2等视频异常事件数据集上进行实验。在CUHK Avenue数据集中,本文方法的AUC(area under curve)为87.2%,比同类方法高2.3%;在UCSD Ped1和UCSD Ped2数据集中,本文方法的AUC值均高于同类其他方法。同时,设计了4个消融实验并对实验结果进行对比分析,本文方法具有更高的AUC值。结论 实验结果表明,本文方法更适合视频异常检测任务,有效提高了异常行为检测任务模型的稳定性和准确率,且采用视频序列帧间运动信息能够显著提升异常行为检测性能。展开更多
To enhance the communication quality of Open Flow controlled all-optical networks,an optical signal-to-noise ratio comprehensive-awareness(OSNR-CA) model based lightpath control scheme is proposed.This approach transf...To enhance the communication quality of Open Flow controlled all-optical networks,an optical signal-to-noise ratio comprehensive-awareness(OSNR-CA) model based lightpath control scheme is proposed.This approach transforms main physical-layer optical impairments into OSNR value,and takes this comprehensive OSNR value of the optical signal along the lightpath into consideration,when establishing the lightpath for the connection request using OpenFlow protocol.Moreover,the proposed scheme makes full advantages of the OSNR monitoring function in each node,and assigns the lightpath according to the comprehensive-OSNR value by extending messages of Open Flow protocol,in order to guarantee the reliable establishment of the lightpath.The simulation results show that the proposed scheme has better performance in terms of packet loss rate and lightpath establishment time.展开更多
文摘人工养殖的雄性梅花鹿在发情期间攻击行为剧增,易造成鹿茸损伤,自动监测其攻击行为能为研究减少攻击行为提供重要依据。本文基于注意力机制和长短记忆序列研究了一种光流注意力网络(Optical flow attention attacking recognition network,OAAR),对梅花鹿的攻击、采食、躺卧、站立行为进行识别。OAAR网络包括前置网络、基础网络和时序网络,前置网络由LK光流算法(Lucas kanade optical flow algorithm)组成,用于提取RGB数据光流信息;基础网络中采用自注意力模块,将ResNet-152网络改造为ARNet152(Attention ResNet-152),用于将RGB、光流数据集经ARNet152提取特征后输入时序网络;时序网络采用添加注意力模块的长短记忆序列(Attention long short term network,ALST),并通过分类器输出行为得分和分类结果。视频数据集包括10942段,共310574帧,划分为攻击、采食、站立和躺卧4个大类,攻击行为又划分为撞击、脚踢和追逐3个小类;训练集、验证集和测试集比例为3∶1∶1。研究结果显示,OAAR模型在测试集上正确率为97.45%、召回率为97.46%、F1值为97.45%,ROC曲线中各类识别效果良好,特征嵌入图中各类行为特征区分度较高,各项结果均优于LSTM、双流I3D和双流ITSN网络,具有较好的泛化能力和抗干扰性。在本研究算法基础上集成的鹿只行为自动识别采集系统,为提高梅花鹿养殖生产管理水平和生产效率提供了技术基础。
文摘目的卷积神经网络广泛应用于目标检测中,视频目标检测的任务是在序列图像中对运动目标进行分类和定位。现有的大部分视频目标检测方法在静态图像目标检测器的基础上,利用视频特有的时间相关性来解决运动目标遮挡、模糊等现象导致的漏检和误检问题。方法本文提出一种双光流网络指导的视频目标检测模型,在两阶段目标检测的框架下,对于不同间距的近邻帧,利用两种不同的光流网络估计光流场进行多帧图像特征融合,对于与当前帧间距较小的近邻帧,利用小位移运动估计的光流网络估计光流场,对于间距较大的近邻帧,利用大位移运动估计的光流网络估计光流场,并在光流的指导下融合多个近邻帧的特征来补偿当前帧的特征。结果实验结果表明,本文模型的m AP(mean average precision)为76.4%,相比于TCN(temporal convolutional networks)模型、TPN+LSTM(tubelet proposal network and long short term memory network)模型、D(&T loss)模型和FGFA(flow-guided feature aggregation)模型分别提高了28.9%、8.0%、0.6%和0.2%。结论本文模型利用视频特有的时间相关性,通过双光流网络能够准确地从近邻帧补偿当前帧的特征,提高了视频目标检测的准确率,较好地解决了视频目标检测中目标漏检和误检的问题。
基金National Key Lab of Broad Band Fiber Transmission and Communication System Technology , Ministry ofEducation
文摘General multi-protocol label switching(GMPLS) based on traffic engineering is one of the possible methods to implement all-optical network. This method implements the network with IP technique and guarantees the quality of service with traffic engineering. Based on the establishment of selecting schemes of optical path and methods of traffic calculation, the wavelength routing algorithm of all-optical network based on traffic engineering is presented by combining with prior route of shortest path and traffic engineering, the algorithm procedures are given, and the actual examples are introduced as well as the analysis on simulation calculation. This research results have certain significance for the achievement of optical switching technique of all-optical network.
文摘目的 视频异常行为检测是当前智能监控技术的研究热点之一,在社会安防领域具有重要应用。如何通过有效地对视频空间维度信息和时间维度信息建模来提高异常检测的精度仍是目前研究的难点。由于结构优势,生成对抗网络目前广泛应用于视频异常检测任务。针对传统生成对抗网络时空特征利用率低和检测效果差等问题,本文提出一种融合门控自注意力机制的生成对抗网络进行视频异常行为检测。方法 在生成对抗网络的生成网络U-net部分引入门控自注意力机制,逐层对采样过程中的特征图进行权重分配,融合U-net网络和门控自注意力机制的性能优势,抑制输入视频帧中与异常检测任务不相关背景区域的特征表达,突出任务中不同目标对象的相关特征表达,更有效地针对时空维度信息进行建模。采用LiteFlownet网络对视频流中的运动信息进行提取,以保证视频序列之间的连续性。同时,加入强度损失函数、梯度损失函数和运动损失函数加强模型检测的稳定性,以实现对视频异常行为的检测。结果 在CUHK(Chinese University of Hong Kong) Avenue、UCSD(University of California, San Diego) Ped1和UCSD Ped2等视频异常事件数据集上进行实验。在CUHK Avenue数据集中,本文方法的AUC(area under curve)为87.2%,比同类方法高2.3%;在UCSD Ped1和UCSD Ped2数据集中,本文方法的AUC值均高于同类其他方法。同时,设计了4个消融实验并对实验结果进行对比分析,本文方法具有更高的AUC值。结论 实验结果表明,本文方法更适合视频异常检测任务,有效提高了异常行为检测任务模型的稳定性和准确率,且采用视频序列帧间运动信息能够显著提升异常行为检测性能。
基金supported by the National High Technical Research and Development Program of China(863 Program)(No.2012AA050804)
文摘To enhance the communication quality of Open Flow controlled all-optical networks,an optical signal-to-noise ratio comprehensive-awareness(OSNR-CA) model based lightpath control scheme is proposed.This approach transforms main physical-layer optical impairments into OSNR value,and takes this comprehensive OSNR value of the optical signal along the lightpath into consideration,when establishing the lightpath for the connection request using OpenFlow protocol.Moreover,the proposed scheme makes full advantages of the OSNR monitoring function in each node,and assigns the lightpath according to the comprehensive-OSNR value by extending messages of Open Flow protocol,in order to guarantee the reliable establishment of the lightpath.The simulation results show that the proposed scheme has better performance in terms of packet loss rate and lightpath establishment time.
