Violence poses an imperative threat to society and public safety and has become a critical issue in recent years. Preventing violent incidents in public places such as roads, streets, playgrounds, and social gatherings is important for maintaining societal well-being and social stability. In the modern time of CCTVs in mega-cities, there is an urgent problem of capturing criminals, leading to a demand for an innovative solution. Our proposed study, introduced FMF-YOLO is based on YOLOv11s, and the main contribution is the integration of the improved Feature Enhancement Module (FEM) and upgraded Feature Fusion Module (FFM) into the main detection architecture. The enhanced FEM is integrated right at the end of the backbone, improving semantic representation and enriching receptive fields by utilizing a multi-branch convolutional architecture integrated with atrous convolutions, thus improving the detection capability of tiny objects without reducing the pixel resolution. The upgraded FFM is integrated into the neck of the base model and is designed upon a modified BiFPN architecture combined with Channel Reweighting and Calibration (CRC) components to efficiently reweight and combine multiscale attributes from various semantic levels with better control. The proposed FMF-YOLO model achieved precision of 97.004% and mAP@0.5 of 93.304% on the K1 dataset. It secured precision of 93.011% and mAP@0.5 of 88.613% on the more challenging K2 dataset. Additionally, we developed an LLM-based system that not only detects violent activities but also generates context conscious incident reports for detected activities.
Saha et al. (Wed,) studied this question.