Video anomaly detection

Motivation

Fall incidents are a major health concern, especially for the elderly population, often resulting in severe injuries and requiring immediate medical attention. Traditional video surveillance systems for fall detection compromise privacy by capturing and processing raw video footage. Our work aims to address this critical challenge by developing a privacy-preserving fall detection system that can accurately identify fall events while maintaining the privacy of individuals in the monitored environment.

Method

Our approach leverages a two-stage pipeline that prioritizes both privacy and accuracy:

1. Pose Estimation: We use OpenPose, an advanced pose estimation technique, to extract skeletal information from video frames. This process transforms raw video data into abstracted human pose representations, effectively eliminating personally identifiable visual information.

2. Spatio-Temporal Graph Neural Networks (ST-GNN): The extracted skeletal data is modeled as a graph, where joints are nodes and their connections are edges. We apply Spatial Temporal Graph Convolutional Networks (ST-GCN) to learn both the spatial relationships between different body parts and the temporal patterns of movement. This architecture is particularly effective at capturing the distinctive motion patterns associated with falls.

This methodology allows our system to detect anomalous events such as falls while discarding the privacy-sensitive visual data from the original video stream.

Results

Our fall detection system demonstrates robust performance across various scenarios. Below is a demonstration comparing the original video footage (left) with the privacy-preserving pose estimation (right) used for fall detection:

The system effectively detects fall incidents while maintaining privacy by operating solely on skeletal data rather than identifiable video content.