Research on Active Detection Methods for Video Anomaly Events Based on Concept Semantics.

Video abnormal event detection is a technology that identifies and detects abnormal events by analyzing video content. Existing methods are mostly limited to manual annotation and visual information, ignoring the unique concepts of abnormal events and the association between concepts. We propose a novel active detection method for video anomaly events based on concept semantics (CS-ISTL) to address these limitations. Specifically, CS-ISTL incorporates semantic similarity-based multi-view representation learning, employing it to transform external knowledge into conceptual semantics effectively. While actively learning to acquire conceptual semantic information, the results are screened through uncertainty measures. In this way, the ability of the model to express abnormal events is improved, and the detection of abnormal video events is realized. Experimental results on the CUHK Avenue dataset and UCSD Pedestrian dataset Experimental results on the CUHK Avenue dataset and UCSD Pedestrian dataset, CS-ISTL improves the AUC indicator by about 10%. Our method significantly outperforms other previously proposed methods because of the characteristics of introducing external knowledge and active learning.