Unsupervised Anomaly Event Detection in Video Based on Autoencoders and Motion Features
Jiangpeng Fu, Yu Zhang, Guofeng Lin, Wentao Fan · 2019
In this article, we propose an unsupervised learning method for anomaly detection in videos based on autoencoders to learn the distribution of normal events. In our model, if the behavior deviates from normal distribution, it is considered as abnormal behavior. We use 2D convolution to extract the spatial feature of the video data. In order to compensate for the temporal information of video data, we incorporate optical flow features which include two consecutive frames of motion features. To improve the expressive ability of our model and fit more complex distribution functions, a `micro network' is constructed and added into the model. Moreover, we add the convolutional long Short-Term Memory (ConvLSTM) in the middle section between encoder and decoder of our model to learn temporal evolution of spatial feature. The effectiveness of the proposed unsupervised anomaly event detection method is validated through experiments that are conducted on four publicly available data sets.