融合注意力机制的自编码器视频人体异常行为检测模型

Autoencoder video human abnormal behavior detection model combined with attention mechanism

在固定监控场景下的人体异常行为检测任务中,行为正常与否判定很大程度上与其发生时的背景息息相关,因此,需要检测模型从全局考虑而非只是行为本身;基于卷积神经网络的自编码器检测模型通常缺乏部分与整体之间的交互,且存在泛化能力过强的不足。针对此两方面问题,提出一种全局自注意力与卷积特征共享的自编码器异常行为检测模型SW-MemAE,该模型以全局注意力捕捉图像整体特征交互信息,在瓶颈处插入记忆模块以约束自编码器对于正常行为的过度泛化,通过视频前四帧和预测帧间的重构误差来判断是否存在异常行为。使用USCD-Ped2、CUHK Avenue两个基准数据集对该模型性能开展实验,结果表明,相比于其他基于预测或重构的视频异常行为检测模型,提出SW-MemAE模型在AUC指标上分别达到95.69%、84.1%,检测性能表现良好。

In the human abnormal behavior detection task in a fixed monitoring scenario, the judgment of whether the behavior is normal or not is closely related to the background when it occurs, so the model needs to detect abnormality from a global perspective rather than the abnormal behavior itself, and the autoencoder detection model based on convolutional neural network usually lacks the interaction between the part and the whole, and the ability of generalization is too strong. Aiming at these two types of problems, a global self-attention and convolutional feature sharing autoencoder abnormal behavior detection model, named SW-MemAE, is proposed in this paper. The global attention captures the overall image feature interaction information, and a memory module is inserted at the bottleneck to constrain the autoencoder the overgeneralization of human normal behavior. At the same time, the model judges whether there is abnormal behavior by the reconstruction error between the first four frames of the video and the predicted frame. The SW-MemAE model has been verified on two benchmark datasets of USCD-Ped2 and CUHK Avenue. The experimental results show that the AUC of proposed model reaches to 95. 69% and 84. 1% which compared with other video anomalous behavior detection models based on prediction or reconstruction, and the performance of the proposed model is better.