Unsupervised Video Anomaly Detection Based on Similarity with Predefined Text Descriptions

Research on video anomaly detection has mainly been based data. However, many real-world cases involve users who can conceive potential normal and abnormal situations within the domain. This domain knowledge be conveniently expressed as text descriptions, such “walking” or “people fighting”, which easily obtained, customized for specific applications, applied to unseen videos not included in training dataset. We explore of using these descriptions with unlabeled datasets. use large language models obtain leverage them detect frames by calculating cosine similarity between input frame CLIP visual model. To enhance performance, we refined CLIP-derived an dataset proposed text-conditional similarity, is a measure two vectors additional learnable parameters triplet loss. The method simple inference process that avoids computationally intensive analyses optical flow multiple frames. experimental results demonstrate outperforms unsupervised methods showing 8% 13% better AUC scores ShanghaiTech UCFcrime datasets, respectively. Although shows −6% −5% than weakly supervised those videos, 17% 5% scores, means comparable require resource-intensive labeling. These outcomes validate detection.