Anomaly Based Camera Prioritization in Large Scale Surveillance Networks

Digital surveillance systems are ubiquitous and continuously generate massive amounts of data, manual monitoring is required in order to recognise human activities public areas. Intelligent that can automatically ide.pngy normal abnormal highly desirable, as these would allow for efficient by selecting only those camera feeds which occurring. This paper proposes an energy-efficient prioritisation framework intelligently adjusts the priority cameras a vast network using feedback from activity recognition system. The proposed system addresses limitations existing three-step framework. In first step, salient frames selected online video stream frame differencing method. A lightweight 3D convolutional neural (3DCNN) architecture applied extract spatio-temporal features second step. Finally, probabilities predicted 3DCNN metadata processed linear threshold gate sigmoid mechanism control camera. performs well compared state-of-the-art violent methods terms large-scale networks. Comprehensive experiments evaluation showed our approach achieved accuracy 98% with F1-score 0.97 on Hockey Fight dataset, 99% 0.98 Violent Crowd dataset.