Detective AI: A Video-Analysis Framework for Shoplifting-Like Event Detection

This paper presents Detective AI, a video-analysis application that combines Florence-2 localization, SegmentAnything Model 2 (SAM 2) masks, and an EvidenceConstrained Object State Transition (ECOST) layer.ECOST follows personitem interactions and evaluatescontainment, persistence, item visibility, and semanticbehavior before raising an event for operator review. Weevaluated its clip-level decision rule on 100 ten-secondUCF-Crime clips, evenly divided between positive andnormal labels. The nal recorded outcomes were 44 truepositives, 45 true negatives, ve false positives, and sixfalse negatives. These counts correspond to 89.00% accuracy, 89.80% precision, 88.00% recall, 90.00% speci-city, and an F1 score of 88.89%. The paper also reviews the recorded notes for all 11 classication errorsand reports the run-time entries in the experiment manifest. Results apply to the evaluated clips and labels.An alert indicates that a clip matched the stated eventrule; it does not establish a person's intent and requireshuman review.