A Robust Deep Learning Approach for Real-Time Violence Detection: Comparative Study of CNN, LSTM, and GRU

Abstract

Time-to-real detection of potential threats has formed the base of maintaining public safety through violence detection systems. The present study dwells on deep learning models that offer improved prediction accuracy for human activity analysis with the changed scope of their AI use. Recent studies use a combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs such as LSTM) in many settings to classify violence in videos obtained from surveillance footage, attaining a substantial drop in the rate of classification errors. This research constitutes a hybrid deep learning framework that combines CNNs for extracting features and LSTMs for analysing violent and non-violent scenes. It was also examined how well the GRUs identified and classified the scenes along with fully connected networks in parallel so as to increase the robustness of the model. CNNs extract spatial features, LSTMs and GRUs capture temporal dynamics. The experiment reaches 92% test accuracy, accomplishing an impressive advance in differentiating violent onset from benign events, achieving superior precision and recall values compared with state-of-the-art benchmarks. This finding delineates the contribution of deep learning in respect of the effective enhancement of automated surveillance, paving the path for safer environments through proactive presence of threat detection.