Seminar: Towards unbounded activity&context awareness in wearables and ubicomp
Dr Alex Casson, Lecturer in Sensing, Imaging and Signal Processing, University of Manchester
Room 1.06, Merchant Ventures Building, Woodland Road, Clifton, Bristol.
Human Activity Recognition (HAR) is a fundamental technology in wearable computing and more generally in ubiquitous computing which enables applications in health/wellbeing, sports, industrial assistance and other domains.
Daniel will describe the work he has done towards more robust, adaptive and scalable HAR. Daniel will show the work on an opportunistic recognition pipeline which dynamically adapts to the sensors available in the user's surrounding. The highlights include the automatic translation of recognition models from one sensor modality to another one, or the brain-guided adaptation of recognition models.
Daniel will show the benefits of Deep Learning for HAR where we showed that wearable datasets are large enough for deep network training and we demonstrated significant performance improvements on benchmark datasets. Daniel will show that transfer of layers in a deep network can be used to reduce training time at equivalent performance. This indicates a path towards a universal and reusable "feature basis" for HAR.
In contrast to high-performance recognition, some of our work also focuses on low-power algorithms which may eventually be part of a sensor front-end in an ASIC, in particular a template matching algorithm running at than 120uW.
Lastly Daniel will touch on our recent work towards "open-ended" activity recognition: systems able to discover and classify activities beyond a pre-defined set, which may play a fundamental role in scenarios of high societal value, such as memory assistants for people with dementia.
This work has been supported by EPSRC, Google FRA, NVidia, Huawei, EU FP7 and others.
For more information please visit the SPHERE seminar page.