Early-Anomaly Prediction in Surveillance Cameras for Security Applications

In the last decade, the number of surveillance cameras has increased significantly, with much research conducted to automate the process of surveillance, as humans cannot manage to monitor all these cameras individually, which may cause errors in public safety or abnormal situations. Also, humans may overlook key details in such abnormal behaviours in surveillance cameras. The proposed approach predicts abnormal behaviour using generative adversarial networks (GANs). GANs are trained using different datasets that contain various behaviours to predict future frames. These future frames are transmitted to a deep learning neural network to classify them as normal or abnormal activities, and future anomalies can be detected before they happen. Our initial results show that depending on the future frames extracted by the GAN model is possible, as these extracted frames either improve the accuracy of the detection model or do not affect it, but they can also be further enhanced to detect more frames at a longer duration and predict anomalies before they happen. Anomalies in surveillance will not only be detected but also predicted before they happen, which will result in the prevention of crimes, reductions in surveillance costs and a safer environment overall.