Human-centric AI security

Abstract

In the field of computer vision, the use of machine learning methods for security involves detection and recognition. When used in conjunction with surveillance, one can enhance the safety it provides. Through human action recognition, unsavoury behaviour can be detected which provides greater peace of mind for the public. In this project, the author created a baseline action recognition framework. Starting with the building of a custom dataset from Closed-Circuit Television (CCTV) footage of an office space. This custom dataset is created with an action recognition model using the human skeletal structures in mind. Therefore, the custom dataset is to only retain videos that contain said structures in order to minimize costs incurred when it is to be sent for manual labelling. Tracking of these skeletal structures is also done in order to properly label the relevant actions recognized with the person doing said action. This is done by repurposing a person re-identification (ReID) framework for tracking of a person within the video. Action recognition is then done using a Spatial-Temporal Graph Convolutional Neural Network (ST-GCN). As an initial test of this framework, the available action classes that were labeled in-house are ‘Running’, ‘Walking’, ‘Standing’ and ‘Sitting’.