Real time anomalies detection in crowd using convolutional long short-term memory network

Violence is a critical social problem and demands to evaluate through computer vision approaches. At present, the incidences of violent actions get grown in the community, particularly in public places due to several economic and social causes. Moreover, our society's populations are increasing day by day and it is challenging to keep citizens within limits as well as monitoring human activities in crowd is too hard. Thus, government organizations including local bodies, require examining such occurrences through smart surveillance. In this research, a lightweight computational architecture has been presented to classify non-violent and violent activities. A model has been proposed to extract time-based features using smart devices, high-speed wireless networks and cloud servers to classify real-time human activities. For this purpose, a deep learning-based model is employed to detect violent activities and assist the stakeholders in exposing such activities in real-time. Convolutional long short-term memory (Conv-LSTM) is employed to extend fully connected LSTM (FC-LSTM) to capture the frame and detect violent actions. The proposed model accomplished 95.16% validation accuracy using a standard crowd anomaly dataset.