Deep Convolutional Long Short-Term Memory Network based video abnormal behavior prediction
Wenqing Mao, Yepeng Guan · 2020
A deep convolutional long short-term memory network (DC-LSTM) based video abnormal behavior prediction method has been proposed. The DC-LSTM deepens the single-layer convolution of the convolutional LSTM, which can be applied to suppress structural information ambiguity in predicting multiple frames. Mean square error is combined with structural similarity to perform loss reconstruction for correlative calculating between predicted and input frames. An adaptive dynamic double-threshold strategy based on the loss reconstruction is adopted to determine whether the video is abnormal. Comparisons with some state-of-the-arts have highlighted the superior performance of the proposal.