Prototype-Guided Contrastive Learning for Unsupervised Video Anomaly Detection with Robust Temporal Scoring

Automatic video anomaly detection remains challenging because abnormal events are infrequent, visually heterogeneous, and weakly bounded in time. This study proposes an unsupervised framework trained only with normal video segments. The framework integrates sliding-window segment construction, dual-view perturbation, a two-branch spatio-temporal encoder, exponential moving-average prototype updating, prototype-guided contrastive optimization, and a robust anomaly score composed of prototype deviation, second-order temporal residual, and local-neighborhood sparsity. Experiments were conducted on UCSD Ped2, CUHK Avenue, and ShanghaiTech under the same input size, segment length, optimizer, and threshold protocol. The proposed model achieved AUC values of 97.4%, 91.8%, and 83.7% on the three datasets, respectively, with an average AUC of 91.0% and an average F1 score of 88.1%. Relative to the baseline contrastive model, the average AUC increased by 2.4 percentage points, and the average F1 score increased by 2.8 percentage points. Across three independent runs, the improvement over the contrastive baseline was statistically significant (paired two-sided t-test, p = 0.018). Ablation and sensitivity analyses indicate that the performance gain is mainly attributable to spatio-temporal joint encoding, prototype traction, temporal residual scoring, and local-neighborhood support. These results show that contrastive representation learning, explicit prototype updating, and temporal-aware scoring can jointly produce a stable representation of normal behavior without using abnormal samples during training.