基于贝叶斯融合的时空流异常行为检测模型

针对直接利用卷积自编码网络未考虑视频时间信息的问题，该文提出基于贝叶斯融合的时空流异常行为检测模型。空间流模型采用卷积自编码网络对视频单帧进行重构，时间流模型采用卷积长短期记忆(LSTM)编码-解码网络对短期光流序列进行重构。接着，分别计算空间流模型和时间流模型下每帧的重构误差，设计自适应阈值对重构误差图进行二值化，并基于贝叶斯准则对空间流和时间流下的重构误差进行融合，得到融合重构误差图，并在此基础上进行异常行为判断。实验结果表明，该算法在UCSD和Avenue视频库上的检测效果优于现有异常检测算法。

Spatial-temporal Stream Anomaly Detection Based on Bayesian Fusion

Focusing on the problem that convolutional auto-encoder network based anomaly detection ignores time information, a novel anomaly detection model based on Bayesian fusion of spatial-temporal stream is proposed. A convolution auto-encoder network is used in spatial stream model to reconstructs video frames, and a convolutional Long Short-Term Memory (LSTM) encoder-decoder network is used to reconstruct short-term optical sequence in the temporal stream model. Then, the reconstruction errors under spatial and temporal stream are calculated separately. Meanwhile, an adaptive thresholds is designed to obtain the reconstruction binary error maps. Finally, the Bayesian fusion strategy is developed to combine the reconstruction error of spatial and temporal stream to obtain the final fusion reconstruction error map based on which the abnormal behavior can be determined. Experimental results show that the proposed algorithm is superior to the existing anomaly detection algorithms in UCSD and Avenue datasets.