基于深度时空卷积神经网络的人群异常行为检测和定位

针对公共场合人群异常行为检测准确率不高和训练样本缺乏的问题，提出一种基于深度时空卷积神经网络的人群异常行为检测和定位的方法。首先针对监控视频中人群行为的特点，综合利用静态图像的空间特征和前后帧的时间特征，将二维卷积扩展到三维空间，设计面向人群异常行为检测和定位的深度时空卷积神经网络；为了定位人群异常行为，将视频分成若干子区域，获取视频的子区域时空数据样本，然后将数据样本输入设计的深度时空卷积神经网络进行训练和分类，实现人群异常行为的检测与定位。同时，为了解决深度时空卷积神经网络训练时样本数量不足的问题，设计一种迁移学习的方法，利用样本数量多的数据集预训练网络，然后在待测试的数据集中进行微调和优化网络模型。实验结果表明，该方法在UCSD和subway公开数据集上的检测准确率分别达到了99%和93%以上。

Abnormal crowd behavior detection and localization based on deep spatial-temporal convolutional neural networks

To handle the issues of low accuracy and lacking training samples in abnormal crowd behavior detection in public places, this paper proposed a method based on deep spatial-temporal convolutional neural networks. In view of the characteristics of crowd behavior in monitoring videos, it first designed a deep spatial-temporal convolution neural network for detecting abnormal crowd behavior by extending 2D convolution to the 3D space according to spatial features of static images and temporal features between the frames before and after the current frame. To locating abnormal crowd, this paper divided video frames into a number of subregions that obtained spatial-temporal samples. Then, it input the samples into the designed deep spatial-temporal convolutional neural network for training and classification, whose results were used to detect and locate abnormal crowd. In the meanwhile, this paper utilized a transfer learning method to deal with the issue of lacking training samples when training the deep spatial-temporal convolutional neural network, where datasets with more training samples were used to pre-train the network which was fine-tuned and optimized on testing datasets with fewer samples. Experimental results show that the detection accuracies on UCSD and subway open datasets are greater than 99% and 93% respectively.