Lightweight AI Spots Violence in Surveillance Video at 130 Frames Per

Researchers in Algeria developed a lightweight AI system to detect violent activity in real-time surveillance video, aiming to close gaps caused by overburdened human operators. Published in Multimedia Tools and Applications, the work reports 90% detection accuracy while processing up to 130 frames per second on live-style inputs. The team, led by Maya Houacine of the University of Science and Technology Houari Boumediene in Algiers, with Khadidja Belattar and Naila Aziza Houacine, says many prior studies used computationally heavy models that can’t run on camera hardware, or they tested on staged footage without validating real-time performance. The pipeline first extracts keyframes to remove redundancy, then evaluates three deep learning architectures, including an enhanced MobileNet variant.