Privacy-Preserving Video Anomaly Detection: A Survey.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Video anomaly detection (VAD) is crucial for public safety but raises privacy concerns due to personally identifiable information in surveillance footage.
• Existing VAD systems often lack transparency, limiting real-world applications and public trust.
• Privacy-preserving VAD (P2VAD) has emerged as a research hotspot, yet current studies are fragmented and often overlook privacy leakage and appearance bias.

Purpose of the Study:

• To systematically review the progress of privacy-preserving video anomaly detection (P2VAD).
• To define the scope and provide an intuitive taxonomy for P2VAD research.
• To identify challenges and opportunities for future P2VAD development and deployment.

Main Methods:

• Comprehensive literature review of P2VAD research.
• Development of a novel taxonomy for classifying P2VAD approaches.
• Analysis of assumptions, learning frameworks, and optimization objectives of various P2VAD methods.
• Evaluation of strengths, weaknesses, and potential correlations among different approaches.

Main Results:

• This article presents the first systematic review of P2VAD, offering a structured overview of the field.
• A clear taxonomy is provided, categorizing P2VAD methods based on their approaches to privacy preservation.
• Strengths, weaknesses, and interconnections of various P2VAD techniques are analyzed.

Conclusions:

• The review highlights the fragmented nature of current P2VAD research and identifies gaps, particularly concerning RGB sequence-based methods.
• Open-access resources, including datasets and code, are provided to facilitate further research.
• Key challenges and future opportunities are discussed to guide advancements in AI development and P2VAD deployment.