Algiers Researchers Achieve 90% Violence Detection with Lightweight AI
An Algerian research team, led by Maya Houacine of the Laboratory of Research in Artificial Intelligence at the University of Science and Technology Houari Boumediene in Algiers, alongside Khadidja Belattar and Naila Aziza Houacine, has developed a lightweight artificial intelligence system designed for real-time violence detection in surveillance footage. This innovative system achieves a 90 percent detection accuracy while processing video at speeds of up to 130 frames per second. The work, recently published in the journal Multimedia Tools and Applications, represents a significant advancement in security technology by addressing the challenge of human operators struggling to monitor vast amounts of live video feeds.
The team specifically focused on creating a system that could operate on real-world hardware, overcoming limitations of existing methods that often rely on computationally heavy models or are only tested on staged footage. Their approach prioritized real-time throughput from the outset, incorporating keyframe extraction techniques to reduce video redundancy and selecting informative frames for analysis. After evaluating various deep learning architectures, the MobileNet-based model proved most effective, balancing 90 percent accuracy with 91 percent precision and 90 percent recall, crucial metrics for minimizing false alarms and missed incidents in security applications. This robust performance was validated against challenging public benchmarks, including RWF-2000 and UCF-Crime datasets.
The system's high processing speed of 130 frames per second significantly surpasses the typical 25 to 30 frames per second required for live video, providing ample headroom for integration into existing surveillance platforms. This capability ensures that alerts are generated in real-time, allowing for timely responses to incidents. The researchers emphasized ethical considerations throughout their work, utilizing exclusively publicly available datasets and ensuring no attempt was made to identify individuals. All data handling adhered to privacy and data protection regulations, reflecting principles of responsible data usage and respect for individual privacy, positioning this Algerian innovation as a practical and ethically sound solution for public safety infrastructure.
