一种改进R(2+1)D网络的暴力行为检测方法

Improved violent behavior detection method for the R(2+1)D network

公共安全中复杂的暴力行为检测具有重要的研究价值。传统的研究方法主要基于手工设计的特征,泛化能力较差,现有的深度学习网络模型泛化能力强但准确率较低。针对上述问题,提出了一个结合R(2+1)D改进网络和密集连接思想的暴力行为检测方法。由于原R(2+1)D残差模块支路中的步长为2的卷积操作忽略了特征图的3/4,所以将其优化为池化操作和步长为1的卷积操作。本实验的数据集共有1500个视频样本,具体包括曲棍球比赛数据集和自制数据集。实验结果证明,改进后R(2+1)D网络相比原网络准确率分别提高了约2.30%和1.00%。另外,引入密集连接思想,将残差模块中的不同卷积层级间建立连接,使残差块中的卷积层输出特征图可重复使用,这在一定程度上减轻了训练过程中梯度消散的问题。通过在相同数据集上进行测试,发现改进后(2+1)D网络相比传统的方法,检测精度进一步提升了约1.47%和0.93%。因此,在公开的经典暴力行为检测数据集上的实验证明,相对于传统的3种网络学习方法,该算法能够更好地表示暴力行为信息,是一种更加简单有效的暴力行为检测方法。

In public security,the complex violence behavior detection is of important research value.Traditional methods are mainly based on hand-crafted features,but they have a limited generalization ability.The existing deep learning network models have a better tolerance,but those kinds of methods face the challenge of obtaining a high accuracy.To solve the above problems,a novel violent behavior detection method is proposed in this paper by combining an improved R(2+1)D network and the dense connection idea.In the branch of the traditional R(2+1)D residual module,the three-quarters of the feature map is ignored due to the convolution operation with strides of 2.In this paper,the convolution operation is optimized for a pooling operation and a convolution operation with strides of 1,and the detection accuracy is increased by 2.3%.Besides,the dense connection idea is adopted into the residual module to establish the connection between different convolutional layers.The improvement could alleviate the problem of gradient dissipation during the training process,and the detection accuracy is further improved by 1.46%.Experimental results on one public dataset and one self-built dataset demonstrate the effectiveness of the proposed method for the complex violence behavior detection.