Effective Deep-Learning-Based Depth Data Analysis on Low-Power Hardware for Supporting Elderly Care

Christopher Pramerdorfer, Rainer Planinc, Martin Kampel; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 394-395

We present a detailed technical insight into a commercial vision-based sensor for monitoring residents in elderly care facilities and alerting caretakers in case of dangerous situations such as falls or residents not returning to their beds during nighttime. We focus on aspects that enable deep-learning-based object classification in realtime on low-end ARM-based hardware, which is prerequisite for a solution that is performant yet affordable, low-power, and unobtrusive. To this end, we introduce an efficient vision pipeline that maps the input depth data to concise virtual top-views. These views are then processed by a set of convolutional neural networks, with a scheduler selecting the most appropriate one based on the current operating conditions and available hardware resources. In order to overcome the challenge of acquiring large amounts of training data in this privacy-critical environment, we pretrain these networks on a large set of synthetic depth data. These concepts are general and applicable to similar vision tasks.