Anomaly Detection from Videos using Graph Convolution Networks

Date

2021-12

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Hundreds of thousands of hours of video are recorded by surveillance cameras every day. Although much object detection and person detection and even anomaly detection is carried out on these video feeds, the methods used have still been fairly traditional and repetitive: from bottom-up approaches using low-level features or leveraging the advent of convolutional neural networks or, more recently, image transformers. We propose a novel semi-supervised learning method to detect anomalies from a pedestrian dataset by representing each frame in the video feed as a graph. We create a graph embedding from video frames, where objects are treated as nodes and handcrafted features between the objects are treated as edges. This embedding is then combined with convolutional features to detect anomalies.

Description

Keywords

graph convolutional network, graph network, graph embedding, anomaly detection

Citation