RH-DATM: generic object tracking and distributed attention-based BiLSTM for violent activity detection

In today’s rapidly evolving digital landscape, the rise of multimedia content sharing has led to a high demand for automated systems that identify and mitigate the dissemination of violent content. The existing research for violent activity detection and reorganization faces some major issues of redundancy and motion discontinuity, undermining the overall effectiveness of violence detection. To tackle these challenges the research presented a Raptor hunt search algorithm enabled distributed attention-based Bidirectional Long Short Term Memory (RH-DATM) model. The distributed attention BiLSTM framework is designed to capture long-term dependencies in temporal sequences, enabling the model to effectively discern violent activities in video data. Temporal ternary patterns are employed to encode intricate temporal dynamics, enhancing the model’s ability to recognize nuanced violent behavior. Object tracking is facilitated through a combination of Generic Object Tracking using Regression Networks which improves the reorganization capability of the model. To optimize keyframe selection, a meticulous process is introduced to identify crucial frames that encapsulate the essence of violent events. This method contributes to the reduction of computational complexity while preserving the integrity of the surveillance footage. The proposed optimization algorithm inspired by the cruise and hunting characteristics fine-tuning the model’s tunable parameters also improves the convergence and overall performance. With an accuracy of 94.67%, sensitivity of 95.36%, and specificity of 95.64% for TP 90, the RH-DATM model outperforms the traditional methods.