基于行为关键语句特征的停车场异常行为识别方法  

Parking Anomaly Behavior Recognition Method Based on Key Sentence of Behavior Sequence Features

作　　者：汪鸿年 苏菡[1,2] 龙刚 王雁飞 尹宽 WANG Hong-nian;SU Han;LONG Gang;WANG Yan-fei;YIN Kuan(School of Computer Science,Sichuan Normal University,Chengdu 610101,China;Visual Computing and Virtual Reality Key Laboratory of Sichuan Province,Chengdu 610066,China)

机构地区：[1]四川师范大学计算机科学学院,成都610101 [2]可视化计算与虚拟现实四川省重点实验室,成都610066

出　　处：《计算机科学》2019年第10期299-306,共8页Computer Science

基　　金：国家自然科学基金(61403266,61403196);人社部留学回国人员科技活动择优项目重点项目;四川省可视化与虚拟现实重点实验室项目(KJ201419);成都大熊猫繁育研究基地项目(CPB2018-02)资助

摘　　要：随着技术的发展和摄像头的普及,人们对智能视频监控的需求越来越高,其中异常行为识别是智能监控系统的关键部分,对维护社会安全有着重要的作用。针对视频数据的时空特性,文中提出了将行为表示为具有时间序列性的关键语句的方法,并将这些关键语句称为行为关键语句。通过对行为关键语句的学习,实现了对停车场场景的异常行为识别。首先,对行为图像序列进行分割,提取前景目标并计算前景目标的运动周期曲线;然后,依据运动周期曲线采用动态时间规整(Dynamic Time Warping,DTW)的方法提取行为关键帧;最后,基于自然语言处理领域中的语义理解的方法,将行为关键帧表征为一系列行为关键语句进行识别。针对关键语句的时序性,采用擅长处理时序数据的长短时记忆神经网络(Long Short-Term Memory Network,LSTM)对行为关键语句进行分类。此外,为解决现有的数据不平衡问题,采用生成对抗网络(Generative Adversarial Networks,GAN)等方法扩充训练集,以增大样本空间,平衡不同类别数据量的差异。在中国科学院CASIA行为数据库和自建行为数据库上的验证结果表明,所提方法对异常行为的平均识别率达到了97%,相比于以前的方法有了明显的提升,证明了行为关键语句能更好地表征行为信息且LSTM模型更适用于学习时序数据背后的模式,因此该方法在停车场场景的异常行为识别任务上具有有效性。With the development of technology and the popularity of cameras,people’s demands on intelligent video surveillance are increasing.Anomaly behavior recognition is a key part of intelligent monitoring systems and plays an important role in maintaining social security.Aiming at the spatio-temporal feature of video data,this paper proposed a method of characterizing behavior as a key sentence with time series,termed Key Sentence of Behavior Sequence (KSBS),and realized the anomaly behavior recognition in the parking scenes by learning key sentences of behaviors.Firstly,the motion sequence is segmented,the foreground target is extracted,and the Motion Period Curve (MPC) of the foreground target is calculated.Then,according to the motion cycle curve,the MPC and DTW method are used to extract the behavior key frames.Finally,based on the semantic understanding method in the field of natural language proces- sing ,the behavior key frames are characterized as a series of behavior key sentence.In light of time series features of key sentences,LSTM,which is expert in dealing with time series data,is used to classify the key statements of behaviors.In order to solve the existing data imbalance problem,GAN is used to expand the training set,thus increasing the sample space and balancing the difference between different types of data.Validation results on CASIA behavior database and self-built behavior database show that the average recognition rate of the proposed method for anomaly behavior is 97%.It is proved that the Key Sentece of Behavior Sequence can better represent the behavior information and the LSTM model is more suitable for learning the patterns behind the time series data,verifying the effectiveness of the proposed method on anomaly behavior recognition in parking scenes.

关 键 词：异常行为识别 深度学习特征 动态时间规整 生成对抗网络 长短时记忆神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]