Anomaly Detection by Predicting Future Frames using Convolutional LSTM in Video Surveillance

Events that do not confront normal behavior are called anomalies and they are extremely arduous to recognize. The recent approaches that deploy a reconstruction approach for anomaly detection, predominantly emphasize minimizing the reconstruction error of training data. These techniques cannot assure larger reconstruction errors in the event of an anomaly. In our work, we propose to systematize the issue of abnormal event detection within a regime of future frame prediction. Provided a set of input video frames $i_1, i_2, i_3 \ldots i_n$, our next-frame prediction model predicts a new frame $i_{n+1}$ instead of reconstructing the same frame $i_{n+1}$. By extending the contributions of Convolutional Neural Networks (CNNs) and Long Short Term Memory (LSTM), we propose ConvolutionalLSTM (C-LSTM) as a predictor to predict the next frame. To scrutinize the capability of the prediction model, we determine the intensity loss between the actual frame and the predicted frame. The larger error between the predicted frame and ground truth facilitates the detection of anomalous events that do not confront the expectation. This paper mainly emphasizes how well the model predicts the future frame and provides a new baseline for abnormal event detection.