Abnormal Activity Detection using CNNs with TimeDistributed Layers

Human Behaviour Recognition is a challenging task that, when correctly achieved, significantly improves people safety and prevention of conflicts. This paper proposes a complete pipeline for abnormal activity detection, built around a Convolutional Neural Network with TimeDistributed layers, designed to analyse sequential data. The results show an encouraging accuracy of 74.69%, obtained by combining two datasets, with a total of 800 video sequences, distributed into 8 categories. Moreover, if certain violent activities are detected, an alert is activated.