Abstract
Video anomaly detection remains challenging because abnormal events are scarce, context-dependent, and operationally sensitive. The objective of this chapter is to develop and evaluate an unsupervised TDA-based framework for modeling surveillance windows without relying on reconstruction error. Trajectories are encoded as weighted spatio-temporal point clouds. Vietoris–Rips persistent homology yields H0 and H1 signatures, persistence images, Betti curves, and interpretable scalar descriptors. Models trained solely on normal data are evaluated across six controlled event classes. Isolation Forest achieved a ROC-AUC of 0.967 and average precision of 0.959, while the Mapper graph separated normal regions from anomaly-dense nodes. Results show that global connectivity and loop structure provide stable, privacy-aware support for explainable surveillance analytics and establish a reproducible bridge between mathematical topology and deployable computer vision.

This publication has 25 references indexed in Scilit: