Abnormal event detection by learning spatiotemporal features in videos

Abstract

Abnormal event detection from video surveillance is a key issue for social security. At present, the challenge lies in the effective feature extraction of video data. In order to solve the problem, a deep learning method based on convolutional autoencoder was proposed in this paper. Firstly, video data are preprocessed to obtain video volumes for subsequent training. Secondly, the video volumes are put into the convolutional autoencoder to learn the spatiotemporal features. Specifically, the model can capture spatial features by performing convolution and learn temporal features by Long Short-Term Memory (LSTM). Finally, abnormal event detection is carried out according to the normalized reconstruction error, which is adopted as the index of anomaly degree. Experimental results show that the proposed method had higher accuracy and generalization ability on the challenging Avenue and UCSD datasets.