Ambient Intelligence from Complex Sensors using Deep Neural Networks and Extensive GPU Training

点击翻译按钮获取中文摘要

The Canadian population is aging similarly to much of the Western world. These demographic changes generate significant societal challenges that are well-documented and known. Among these challenges, one of them is to maintain semi-autonomous populations and the elderly at home for as long as possible. Delaying their institutionalization is a more humane solution and is generally less costly for society. However, in that regard, support from caregivers or healthcare professionals who make home visits becomes crucial. In order to facilitate the task of these persons, smart homes have emerged in recent decades as a relevant and inexpensive alternative to monitor the target people's activities and ensure their safety. Although several teams around the world, including ours, are now exploiting these technologies in real residences (compared to laboratories), there are still many challenges to achieve their full potential. Indeed, most smart homes rely on simple ambient technologies such as motion detectors, electromagnetic contacts, and environmental sensors. These types of sensors only provide a crude and imprecise picture of the activities going on between the walls. In order to take activity recognition to the next level, it becomes imperative to exploit sensors that are more expressive in terms of information quality. For example, our team has used ultra-wideband radars and thermal cameras in recent projects. These technologies allow more complex data to be collected, but still protect the privacy of residents because the data is not identifying. However, to fully take advantage of these technologies, it becomes necessary to use complex deep learning models. Although many simple architectures are easily trainable on standard computers with good graphics cards, state-of-the-art models require a lot of computing power and specialized hardware to properly investigate them. Thus, our team is currently hampered in its scientific progress because it does not have access to the appropriate resources. In this project, we are requesting a budget to acquire a machine dedicated to training deep neural networks equipped with 8 NVIDIA RTX A5000 GPUs. This computer is essential and will have a significant impact on our team, not to mention the huge benefits for the university community and local businesses.