A Skeleton-based View-Invariant Framework for Human Fall Detection in an Elevator

This paper considers the emergency behavior detection problem inside an elevator. As elevators come in different shapes and emergency behavior data are scarce, we propose a skeleton-based view-invariant framework to tackle the camera view angle variation issue and the data collection issue. The proposed emergency fall detection model only needs to be trained for a target camera, which is deployed in an elevator at a manufacture's lab, from which a large amount of training data can be collected. The deployment of a source camera, which is in a customer-side elevator, hence can be customized and almost no training effort is needed. Our framework works in four stages. First, a 2D RGB input image is taken from the source camera and a 2D human skeleton is obtained by 2D pose estimation (AlphaPose). Second, the 2D skeleton is converted to a 3D human skeleton by 3D pose estimation (3D pose baseline). Third, a pre-trained rotation-translation (RT) transform (Procrustes analysis (PA)) aligns the 3D pose representations to the target camera view. Finally, a dual 3D pose baseline deep neural networks (D3PBDNN) model for human fall detection is proposed to perform the recognition task. We gather a human fall detection dataset inside different elevators from various view angles and validate our proposal. Experimental results successfully attain almost equivalent accuracy to that of a source camera-trained model.