Algorithms enabling a robot to learn models of complex human behaviour

This project is part of a large EU funded project (FP7) called Strands, which will involve 30+ people, of whom a team of 6 will work at the Intelligent Robotics Lab at the University of Birmingham.

In the overall project we investigate the task a robot that must perform 4D mapping. By this we mean producing not just a map of space (3D mapping), but also of the activities that occur within that space.
Such activity maps will be essential for future robots since how a robot should act within a space depends in large part on how the humans that also use that space behave. A precise description of the PhD project itself is given below.

The IRLab at Birmingham is a leading European lab working in all aspects of intelligent robotics. The group has five faculty, six research fellows and more than ten research students at any one time. The lab has a state of the art robot lab with access to a wide range of advanced platforms for mobility and manipulation.

In the first instance contact Jeremy Wyatt and Nick Hawes ([Email Address Removed], [Email Address Removed]), with your CV and transcript. Put "Strands PhD application" in the subject line.

Detailed description:

The activity in an environment over a period of time can typically be viewed as being composed of many smaller events and interactions. Many of these events are periodic, e.g. the cycle of meals over a day. There will be many di fferences across repetitions, e.g. in the precise agents and objects involved, but the expectation that activity is to an extent periodic (particularly at the daily and weekly level), provides the challenge of this
task: how to recognise, and exploit this in order to build compact models of behaviour. The work will build on existing work on learning models of activity. The particular aspect to be addressed in this PhD is how the ordering of subevents, during a cycle of activity can be learned, taking account of the fact that sometimes particular events may not occur, or that their ordering may be changed. Where there are large degrees of variation in speci c behaviour during a complex event (e.g. lunch) we will extend the above approach to improve discovery of periodicity. We will model variables such as the numbers of occurrences of event classes in an area, within a time period. To do this we will using Markov process extensions of Poisson processes. Our key contribution relative to existing work will be to extend the model to allow multiple instances of multiple event classes (rather than multiple instances of a single event type) to be generated in any spatio-temporal interval, and for the poisson variables for those event classes to be linked. This approach can then be employed to perform anomaly detection at run time, by calculating the data likelihood for numbers of detected events in a period associated with a meta-event.