MAGNAS: A CCTV Searching Platform for Digital Investigation

Abstract

Closed-circuit television (CCTV) systems generate huge amounts of video material, making manual inspection slow, inefficient, and prone to missing evidence. As surveillance scenarios grow more complex, investigators require automated tools capable of understanding extensive, unstructured recordings and supporting natural-language search, which most existing systems cannot provide. Current video-analysis systems usually focus on isolated tasks such as detection, action recognition, or captioning, producing fragmented outputs that lack the consistency and explainability needed for forensic work. This research presents MAGNAS, an intelligent CCTV retrieval platform that converts raw footage into a searchable metadata index using a multi-phase pipeline integrating person detection, multi-object tracking, visual attribute extraction, and action recognition. It generates structured representations including bounding boxes, timestamps, appearance attributes, and actions stored in an SQLite database. Users describe people or events via a natural-language interface, and the system translates these into structured filters for accurate set-theoretic retrieval. MAGNAS was evaluated using mAP for detection, Recall@K and Precision@K for retrieval, and temporal IoU for action alignment. Results show strong person-detection performance and high retrieval accuracy, especially for queries with multiple appearance attributes. Attribute extraction was largely precise, despite challenges with delayed VLM processing, inconsistent action detection under difficult camera angles, and limited temporal accuracy (tIoU = 0.2649). Overall, MAGNAS significantly reduces search time and improves identification of people of interest in large-scale CCTV archives, highlighting the value of organized, explainable video indexing for future investigative technologies and law enforcement applications.