South Dakota State University
Fast and Sustainable Video Anomaly Detection with Continual Learning
Abstract
dc:description.abstractReal-time video anomaly detection systems deployed in surveillance, healthcare, and industrial environments face continuous distribution shifts in lighting, viewpoint, and activity patterns. Existing models often experience performance degradation under these conditions and may suffer catastrophic forgetting when adapting to new environments. This thesis proposes RegiGrow, a parameter-efficient continual adaptation framework built on the Flashback retrieval pipeline. RegiGrow integrates Mixture-of-Experts Low-Rank Adaptation into a frozen ImageBind encoder, enabling sequential domain adaptation without modifying the pretrained backbone. A lightweight router maps visual regime features to a distribution over LoRA experts, each specializing in a distinct normal operating regime. The central contribution is an entropy-triggered expert growth mechanism that dynamically expands model capacity in response to detected distribution shift. Catastrophic forgetting is mitigated through a class-balanced experience replay buffer preserving representative segments from prior domains. RegiGrow is evaluated under a continual learning protocol across UCF-Crime, ShanghaiTech, and UBnormal, with sequential training and cross-phase evaluation. The system achieves a final average AUROC of 0.6678, demonstrating positive backward transfer (+0.0136) and strong forward transfer to unseen domains (+0.1119 on ShanghaiTech, +0.0727 on UBnormal) without catastrophic forgetting.
Degree
thesis:*- Name thesis:degree_name
- Master of Science (MS)
- Level thesis:degree_level
- Thesis - Open Access
- Discipline thesis:degree_discipline
- Electrical Engineering and Computer Science
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Amasa, Preethi
- Contributors dc:contributor
-
- Chulwoo Pack
Subjects
dc:subject × 1Rights
dc:rightsIdentifiers
dc:identifier.*- Repository record dc:identifier
- https://openprairie.sdstate.edu/etd2/2021