Dual-Stage Deep Learning Framework for Effective Public Physical Violence Detection
Adusumilli Divya, D. Sree Lakshmi, P. Niveditha, Pabolu Siva Naganjana Sreya Sri, Venigalla Rohith, Venkat Bhavan Tati · 2024
Violent acts endanger both personal safety and social peace. Various solutions have been considered to minimise such behaviours, including the establishment of surveillance systems. The development of surveillance equipment that can detect aggressive behaviour autonomously is critical. The suggested system requires two stages to be executed. It must first detect the presence of persons inside video frames, followed by the extraction of frames containing potential violence. Unnecessary frames are then removed, leaving only pertinent instances of aggressive behaviour that are preserved as separate pictures. Whenever feasible, the technology recognises the faces of persons taking part in the activities.. Two pre-trained models are combined for higher accuracy and faster processing, addressing limitations of relying solely on Convolutional Neural Networks. This innovative system holds promise for enhancing safety and security. This two-staged approach comprises Faster RCNN Inception V2 COCO and MobileNet V2 pre-trained models that offers improved accuracy and serves as a foundation for constructing the full model.