Doors and Corners

VEGA: Video Evidence-Grounded Analysis VEGA is an AI-powered video anomaly detection and investigation system designed to discover unusual events, understand their context, and explain its findings through verifiable visual evidence. Traditional surveillance systems often depend on predefined event categories, such as fires, accidents, or suspicious activities. VEGA takes a different approach. It looks for abnormalities even when it cannot recognize or name what happened. The system examines video through multiple complementary methods, including optical flow, spatial statistics, object detection, segmentation, and temporal analysis. It maintains camera-specific historical context to distinguish ordinary activity from unusual behavior, including persistent anomalies that may not produce sudden changes. Rather than treating individual model predictions as definitive conclusions, VEGA combines observations into evidence-grounded incident hypotheses. Its investigation pipeline uses Meta's Muse Spark 1.3 to examine video alongside selected visual evidence, reason about possible explanations, and identify uncertainty. Meta's SAM 3.1 supports spatial investigation by segmenting relevant objects and regions. VEGA also provides an interactive interface where users can inspect detected events, replay source footage, examine supporting evidence, and review AI-generated interpretations. The central idea is simple: an anomaly detection system shouldn't need to know what something is to notice that something is wrong. By combining class-free anomaly discovery, persistent scene memory, specialist computer vision, and multimodal reasoning, VEGA explores a more general approach to video understanding that prioritizes discovering unexpected phenomena, preserving evidence, and investigating them rather than forcing every observation into a predefined category.