Surveillance and non-surveillance violence detection using exponential local attention module–based Elman recurrent neural network

Violence detection is critical for ensuring public safety, as it relies on accurately analysing video data from surveillance footage. However, existing detection algorithms have limitations, which often struggle to identify complex motions and low-intensity actions in non-surveillance videos. To overcome this limitation, an Exponential Local Attention Module–based Elman Recurrent Neural Network (ELAM-ERNN) is proposed for accurate violence detection using both surveillance and non-surveillance videos. Initially, video clips are collected from three distinct datasets and key frames are extracted by preprocessing phase. The proposed ELAM-ERNN model detects violence by using 3 Dimensional – Convolutional Neural Network (3D-CNN) extracted spatio-temporal features to distinguish normal from unusual activities, even in unstructured environments with dynamic camera movements. The proposed ELAM technique enhances detection performance by dynamically focusing on high-motion temporal regions to identify subtle violence, while the ERNN captures sequential patterns to learn the temporal coherence of violent actions across multiple frames. Experimental results of the proposed ELAM-ERNN model achieve an accuracy of 99.89% and 99.56% for the Hockey Fight and Violent Flow datasets, which are higher than the existing approach, namely RA-ConvLSTM.