New AI Method Detects Violence in CCTV Footage with 95% Accuracy Using Short Video Clips

Researchers have developed a novel approach to spotting violent behavior in security camera feeds by processing footage in short bursts rather than analyzing entire videos at once. The Short-Window Sliding Learning framework segments CCTV recordings into 1-2 second clips, allowing the system to capture rapid, aggressive actions without losing temporal detail.

How the Method Works#

The approach diverges from traditional violence detection models that train on lengthy video sequences. Instead, each brief clip is processed using LLM-based auto-captioning to generate detailed labels automatically—eliminating the need for expensive manual annotation. By preserving all frames within these short windows, the system maintains the temporal flow essential for recognizing sudden, explosive movements that characterize violent acts.

Accuracy and Real-World Performance#

The framework achieved 95.25% accuracy on the RWF-2000 violence detection benchmark, a standard dataset in the field. When tested on longer surveillance videos from the UCF-Crime dataset, the method reached 83.25% accuracy, suggesting strong generalization beyond its training conditions. These results indicate the approach works reliably whether analyzing short clips or extended footage.

Practical Applications#

The method's design makes it particularly suited for live surveillance systems. The short-window approach enables rapid processing, critical for security operations that need to identify threats as they unfold. By automating the labeling process through language models, the framework could significantly reduce the time and expertise required to build violence detection datasets for new surveillance environments.