治安监控视频中暴力行为的识别与检测

Recognition and Detection of Violence in Public Security Surveillance Video

为了充分提取治安监控视频中的时空特征和时序特征,并对暴力行为进行准确的识别与检测,提出一种基于三维卷积神经网络(3DCNN)和卷积长短期记忆网络(ConvLSTM)的暴力行为识别算法。首先,采用一种通用视频描述符—3DCNN结构,提取视频的短时特征,这些特征封装了视频中与目标和场景相关的背景信息,然后,构建ConvLSTM网络对3DCNN提取的短时特征在时间轴上进行建模,进而充分提取视频的高层时序特征。最后,利用Sigmoid函数分类行为动作。为了验证该算法的高效性,对所提出的方法在暴力行为数据集Hockey上进行验证,达到了98.96%的识别精度。测试结果表明,该融合模型在检测效果上优于目前人工提取特征的方法和深度学习的方法。

In order to fully extract the spatio-temporal and temporal features in the security surveillance video,and accurately identify and detect violent behaviors,a violent approach based on three-dimensional convolutional neural network(3DCNN)and convolutional long and short-term memory network(ConvLSTM)is proposed:Behavior recognition algorithm.First,a general video descriptor—3DCNN structure is used to extract short-term features of the video.These features encapsulate the target,scene,and background information related to the video in the video.Then,the ConvLSTM network is constructed to extract short-term features from 3DCNN in time.Modeling on the axis,and then fully extract the high-level timing features of the video.Finally,the Sigmoid function is used to classify actions.In order to verify the efficiency of the algorithm,the proposed method is verified on the violent behavior data set Hockey,and the recognition accuracy is 98.96%.The test results show that the fusion model is superior to the current manual feature extraction methods and deep learning methods in detection effects.