Video anomaly detection for edge-based IoT systems: A survey of input modalities and real-time applications

• Reviews video anomaly detection methods with a focus on input data modalities. • Analyzes RGB, skeleton, and multimodal approaches under real-time IoT constraints. • Summarizes edge-readiness using latency, model size, FLOPs, and FPS metrics. • Identifies gaps between benchmark accuracy and deployable edge solutions. • Discusses future directions for lightweight and privacy-aware VAD on edge devices. With the vast amount of video data generated daily, researchers have become increasingly interested in extracting meaningful information, particularly for analyzing abnormal events. This growing interest has accelerated progress in video anomaly detection (VAD) as a specialized subfield of computer vision, attracting considerable attention due to its potential applications in real-time scenarios such as elderly care, smart homes, and intelligent surveillance. To provide a comprehensive understanding of this rapidly evolving field, several systematic reviews have been conducted to help new researchers enter the field and assist experienced groups in keeping pace with recent advancements. However, existing surveys lack a focused analysis of how different input data modalities impact the performance of VAD systems, particularly from a privacy-preserving perspective. Understanding the effectiveness of various data modalities and data collection strategies is essential for protecting personal information in computer vision applications. Furthermore, the feasibility of deploying VAD models in real-time Internet of Things (IoT) environments remains underexplored, where low latency, limited resources, and strict privacy requirements are critical considerations. Although edge computing has been increasingly adopted to address these challenges, most studies overlook the deployment of VAD frameworks on resource-constrained devices. Integrating edge-based VAD systems with federated learning algorithms represents a promising direction for enabling privacy-aware and scalable real-world systems. Rather than providing a method-centric summary, this survey reorganizes the VAD literature from a deployment-oriented viewpoint, highlighting how input modality choices fundamentally affect privacy preservation and real-time feasibility on edge-based IoT systems. This work specifically reviews studies published between 2020 and 2025.