A New Automated Threat Detection Framework Using Adaptive and Multi-Head Cross Attention-Based ShuffleNetV2 for Abnormality Classification Along with Object Detection and Tracking Procedures

-With recent advancements in both Artificial Intelligence (AI) and Internet of Things (IoT) skills, it is now easier than ever to create surveillance devices that accurately recoganize persons who may pose a security danger to others in real-time. Certain technologies require manual assessment for the identification of violent or illegal circumstances while recording footage of surveillance cameras is likewise a challenging and undependable activity. So, a novel automatic threat detection technique is implemented with a deep learning mechanism. Initially, essential surveillance videos are taken from the benchmark sources. Next, object detection and tracking process is carried out in collected videos using Transformer-YoloV9 (Trans-YoloV9). Once, the object detection and tracking procedure is completed abnormality classification process is executed. Here, several abnormalities in the object detected and tracking model will be classified using Adaptive and Multihead Cross Attention-based ShuffleNetV2 (AMCA-SNetV2), which helps to detect suspicious objects. Further, the efficiency of abnormality classification is enhanced by tuning several parameters AMCA-SNetV2 using an enhanced Intelligent and Randomized Dollmaker Optimization Algorithm (IR-DOA). Finally, abnormality-classified outcomes are obtained from AMCA-SNetV2 which supports automatic threat detection from the surveillance video. Later, various analyses are validated to observe the efficacy of the offered approach over conventional mechanisms.