Real time violence detection in surveillance videos using Convolutional Neural Networks

Real-time violence detection with the use of surveillance is the process of using live videos to detect violent and irregular behavior. In organizations, they use some potential procedures for recognition the activity in which normal and abnormal activities can be found easily. In this research, multiple key challenges have been oncorporated with the existing work and the proposed work contrast. Firstly, violent objects can’t be defined manually and then the system needs to deal with the uncertainty. The second step is the availability of label dataset because manually annotation video is an expensive and labor-intensive task. There is no such approach for violence detection with low computation and high accuracy in surveillance environments so far. The Convolutional Neural Network’s (CNN) models have been evaluated with the proposed MobileNet model. The MobileNet model has been contrasted with AlexNet, VGG-16, and GoogleNet models. The simulations have been executed using Python from which the accuracy of AlexNet is 88.99 and the loss is 2.480 (%). The accuracy of VGG-16 is 96.49 and loss is 0.1669, the accuracy of GoogleNet is 94.99 and loss is 2.92416 (%). The proposed MobileNet model accuracy is 96.66 and loss is 0.1329 (%). The proposed MobileNet model has shown outstanding performance in the perspective of accuracy, loss, and computation time on the hockey fight dataset.