摘要
针对复杂背景及遮挡等原因引起人群异常行为检测性能低的问题,本文提出了一种基于综合光流特征描述符(SOFD)及轨迹的人群异常行为检测方法。首先,根据人群光流场变化计算人群运动速度、加速度、方向和能量,并利用上述特征设计新的时空特征描述符,即SOFD。其次,利用KLT追踪算法获得人群运动轨迹单帧图。最后,基于所获取的上述特征,设计双流卷积神经网络(TS-CNN)以检测人群异常行为。仿真结果表明:与现有主流算法相比,复杂环境下本文方法具有较高的异常行为检测准确率、较好的泛化性及稳健性。
Focusing on the problem of low detection performance of crowd abnormal behavior conducted by complex background and occlusion as well as other factors,a crowd abnormal behavior detection method is proposed based on synthetic optical flow feature descriptor(SOFD)and trajectory.First,the velocity,acceleration,direction and energy of crowd motion are calculated according to the change of crowd optical flow field.Then,a new space-time feature descriptor,i.e.,SOFD,is constructed based on the above characteristics.Third,the Kanade-Lucas-Tomasi(KLT)tracking algorithm is employed to obtain the single frame of crowd motion trajectory.Finally,the two stream convolution neural network(TS-CNN)is depicted with the abovementioned characteristics to detect the crowd abnormal behavior.Compared with the existing state-of-the-arts algorithms,simulation results show that the proposed method has higher accuracy,better robustness and wide application range in abnormal behavior detection under complex environments.
作者
王洪雁
周梦星
WANG Hong-yan;ZHOU Meng-xing(College of Information Engineering,Dalian University,Dalian 116622,China;School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China)
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2020年第6期2229-2237,共9页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目(61301258,61271379)
中国博士后科学基金项目(2016M590218)
河南省高等学校重点科研项目支持计划项目(14A520079)
河南省科技攻关计划项目(162102210168).
关键词
计算机应用
人群异常行为检测
综合光流特征描述符
轨迹
双流卷积神经网络
computer application
crowd abnormal behavior
synthetic optical flow feature descriptor(SOFD)
track
two stream convolutional neural networks(TS-CNN)
