Violence detection for surveillance systems using lightweight CNN models
N. Suba, Abhinav Verma, Pallavi V. Baviskar, Satishkumar L. Varma · IET conference proceedings. · 2022
Due to significant safety concerns, today, many places are filled with CCTV cameras, but it still uses manual security to overlook these, which leaves much space for human error, be it negligence or a dangerous situation. This may result in many anomalies such as violence, hostage, or fires. To prevent this, an intelligent and automated system has been implemented that tries to overcome this using deep learning techniques. Using Deep Learning models utilizing Convolutional Neural Networks (CNN), specifically the lightweight models - MobileNetV2 and ResNet50V2 are used to identify threat movements in the given frames over three datasets, namely UCF crime, Real life violence situations, and UBI-fights. The violence in the video is detected using frames, and accuracy is measured. A threat is detected by the system from the video frame, based on which the system for the situation generates an alert. In extreme cases, alerts are sent directly to nearby police stations and emergency services with accurate location information while also indicating the suspicious activities at an instance of time.