=== Background ===
The Wearable Computing Laboratory at ETH Zurich (http://www.wearable.ethz.ch/) develops methods to recognize complex and hierarchical human activities from data captured from sensors placed on-body, such as those found in mobile phones.
One goal of our group is to achieve an automatic “life log” of the user’s activities using mobile phones sensors. An example of a life log could be: “yesterday, you were taking a coffee with your friend, then went for shopping, did cook a cake, and watched tv before going to sleep”. This has promising uses to support people with memory loss and dementia. It also enables a lot of exciting applications when combined with social networks, such as the automatic update of the user’s profile on facebook, and pervasive gaming.
In order to realize such a life log the system must be able to recognize location, modes of locomotion, postures, gestures, and infer higher-level activities from these primitives. This is done with streaming signal processing and machine learning techniques. Current approaches follow a “learning by demonstration” principle, where the user is requested to provide training data to the system. This is a major limitation currently, as the size of the training datasets remains small and does not capture the rich variability of human activities.
=== Smart-Days: “Smart Distributed daily living ActivitY-recognition Systems” ===
In a Swiss-funded project which involves ETH Zürich and the University of Applied Sciences atYverdon, we develop a novel crowd-sourcing-based approach to recognize complex human activities. The project time span is 2011-2014.
The Smart-DAYS system is composed of: multi-modal sensor nodes providing data relevant to the user’s activities and capable of local data interpretation (e.g. motion sensors nodes); an on-body mobile device (e.g. phone) that is fusing the sensor node data to infer the user’s activities; and a cloud server backend storing collective activity models.
Smart-Days offers these advances over the state of the art:
* The activity models are obtained from a multitude of users providing sporadic annotations about their current or past activities. Thus a large and rich set of human activities can be captured in a bottom-up process;
* The set of activities to recognize is not statically defined at design-time; it can grow to encompass new activities as they are discovered at run-time;
* The system is capable of adapting activity models at run-time to cope with changing user behavior patterns;
* The system can share and reuse the knowledge acquired between the user’s of the system;
* The system can exploit online data sources to bootstrap the recognition capabilities;
* Essentially, the system will be able to recognizing an unbounded, growable, and adaptable set of human activities in open-ended environments.
In order to realize this, Smart-Days approach combines:
* High-level activity recognition based on unsupervised hierarchical clustering and semisupervised techniques
* Crowdsourcing of activity model acquisition, exploiting the knowledge acquired by other users’s devices and web databases to bootstrap activity recognition.
At ETH Zürich we focus on the development of the crowd-sourcing approach to activity recognition. The University of Applied Sciences at Yverdon focuses on unsupervised hierarchical data clustering techniques. These two approaches are combined into a series of joint evaluations in a large scale deployement. Thus, a tight collaboration between the two institutes is foreseen.
=== Job description ===
We offer a PhD position within the framework of the 3 year long (2011-2014) Smart-Days project. In this position you will be responsible for one of the project’s work package. This work package comprises all elements required to achieve a robust crowd-sourced acquisition of human activity models. You will closely collaborate with the project’s partners throughout the duration of the project.
Your work environment will be multinational with frequent travels to the partner’s location.
Within this project, your research topics will include (but are not limited to):
* Activity and context recognition on mobile phones: Real-time streaming signal processing and machine learning on mobile phones or embedded systems
* Online adaptive machine learning: including subsets of supervised, semi-supervised, unsupervised techniques, as well as transfer learning and multitask learning
* Incentive design and capture of sporadic user feedback: methods allowing to capture user feedback are devised, allowing to maximize information gain and minimize user disturbance.
* Real-world deployment and evaluation: experiments during several weeks or months with a mobile platform deployed with an app store.
* Multimodal data fusion
Starting date: ASAP
=== Requirements ===
The candidate has a diploma, MSc, or equivalent in electrical engineering, micro-engineering, computer science or mathematics.
He has strong interests in mobile computing, machine learning/pattern recognition, signal processing, adaptive and learning systems, and in the combination of theoretical and experimental research.
Fluent spoken and written English is mandatory.
=== Contact and application ===
For further information about the Smart-Days project and your contribution within it, please contact Dr. Daniel Roggen (droggen(at)ife.ee.ethz.ch), or Prof. Andres Perez-Uribe (andres.perez-uribe(at)heig-vd.ch).
If you are interested and believe that you qualify, please send your application to Prof. Gerhard Tröster (troester(at)ife.ee.ethz.ch). Include:
* Curriculum Vitae with the names and contact details of at least 2 references
* a list of exams and grades obtained
* a cover letter explaining how your skills and research interests fit the project
For more information: