The International Association for the Exchange
of Students for Technical Experience



Ref. No:  UK-2019-200-08
Country: United Kingdom
Company information

Employer: Cardiff University
Address: --
Business or product: Computer Science & Informatics
Official responsible: --
Phone number: --
Fax: --
E-mail: --
Website: http://www.cs.cf.ac.uk
Working place: Cardiff
Nearest internat. airport: Cardiff
Nearest public transport: Cardiff Queen St.
Number of employees: 500
Working hours per week: 37.0
Daily working hours: 8.0

Student required

General discipline: Computer and Information Sciences
Field of study: Artificial Intelligence
Study level:
Previous training: No
Language required for training:
English excellent
Other requirements:

Work offered

Kind of work:
Deep learning a “virtual guidedog”

In this project a person wearing a headmounted 3D video camera and an inertial measurement unit (records orientation and acceleration) will walk around city streets while we record data. We will then attempt to predict the data from the inertial sensor from the video data. When the video and inertial data are from the same instance in time then prediction is trivial, for example uniform leftward motion in the camera image would predict that the inertial sensor will report rightward rotation. The interesting question is how far ahead can we predict the output of the inertial sensor from the video image. Or in other words, how far ahead can we predict the behaviour of a person by knowing what they are looking at? For example, if there is an obstacle ahead of the person we would expect the person to turn left or right as they approach the obstacle. Therefore an visible obstacle in the video image would predict a rotation signal from the inertial sensor several seconds later. But what other features in the scene (picked up by the video camera) predict changes in the inertial sensor output (which results from movement of the person)? If we could build a system that successfully predicts future changes in the movement of the person then such a system could be used as a “virtual guidedog” to tell a person with limited vision (due either to a clinical problem with vision or poor visibility) when to stop, change speed or change direction. This project will use first calculate action relevant quantities across the image such as the distance, time-to-contact and trajectory direction of objects within the scene and then apply modern deep learning approaches. The student will have the opportunity to interact with PhD students and postdocs who use deap learning and to work alongside people working on our recent EPSRC funded project on predictive vision systems.
Category: Research and development
Number of weeks offered: min. 8 | max. 10
Within period: from 13.05.2019 to 20.09.2019
Work closed:
Gross pay: 295 GBP per week
Max. deduction to be expected: tax


Lodging will be arranged by: IAESTE and Cardiff University
Canteen facilities available at work: Yes
Estimated cost of lodging: 100 GBP per week
Estimated cost of living incl. lodging: 200 GBP per week

Additional information

Deadline for nomination: 19.04.2019
Not public info:
Additional information: --
Reserved training:- - - - - - - - - -
Postpraxe:postpraxe COBE
IAESTE internal comment:--