一种时序边界注意力循环暴力行为检测神经网络

Temporal Edge Attention Recurrent Neural Network for Violence Detection

摘要:

暴力行为检测是行为识别的一个重要研究方向，在网络信息审查和智能安全领域具有广阔的应用前景.针对目前的时序模型在复杂背景下不能有效提取人体运动特征和常规循环神经网络无法联系输入上下文的问题，本文提出一种时序边界注意力循环神经网络TEAR-Net.首先，以本文提出的一种全新的运动特征提取模块MOE为基础，在保留输入视频段序列背景信息的前提下加强运动边界区域.运动边界对于动作识别的作用要远大于图像其他区域，因此运动边界加强能够有效提高动作特征的提取效率，从而提升后续网络的识别精度.其次，引入了一种全新的结合上下文语境和注意力机制的循环卷积门单元(CSA-ConvGRU)，提取连续帧之间的流特征以及不同帧的独立特征，并关注关键帧，能够极大提升动作识别的效率，以少量参数和较低计算量的代价掌握视频流的全局信息，从而有效提高识别准确率.本文提出的模型在目前最新的公开数据集RWF-2000和RLVS上进行了多种实验.实验结果表明，本文提出的网络在模型规模和检测精度上均优于目前主流的暴力行为识别算法.

Abstract:

Violence detection is one of the most important research topic in behavior recognition，which has great potential applications in network information review and intelligent security.The published works cannot keep their performance in the complexity environments， because they cannot effectively extract movement features and contact consecutive frames. Hence， a novel method is proposed in this paper，which is referred to as temporal edge attention recurrent neural network (TEAR-Net). First，we propose a novel motion object enhancement (MOE) module， which enhances the motion edge while keeping the background information of the video sequences. Because the motion edge has a much greater effect on motion recognition than other areas of the image，the enhancement of motion edge can effectively improve the extraction efficiency of action features，and thus the recognition accuracy is improved. Then we introduce a novel recurrent convolutional gate unit CSA-ConvGRU，which combines context and attention mechanism. It can extract the stream features among consecutive frames and the independent features of each frame. Attention mechanism can help to focus on key frames，which greatly improve the efficiency of action recognition，capture the global information of the video stream with a lower cost， and thus effectively improve recognition accuracy. The proposed model has been tested on the currently lastest public datasets RWF-2000 and RLVS. The experimental results show that the proposed model outperforms the state-of-the-art violence detection algorithms in terms of computational cost and detection accuracy.