Exploring mobile trajectories ...

David Mountain - PhD Thesis
Submitted December 2005

Dept Information Science
City University, London

Animations demonstrating different approaches to prediction

Three approaches to predicting the future location of a moving point object are described below. Each was developed and tested as part of the "Exploring Mobile Trajectories .. " PhD research; the approaches are speed-heading prediction, temporal proximity and spatial proximity. An animation is provided for each approach, showing a sequence of prediction surfaces generated for each point that comprises a particular space-time path.

Speed-heading predictions

Speed-heading prediction surface animation
(1.5mb)

Predictions based upon recently exhibited speed and heading were found to be most effective in situations where movement was reasonably consistent. When predicting into the future, a dilemma is faced about how far into the past it is necessary to look in order to preserve a high degree of diversity within the set, but not include behaviour from the more distant past which is no longer indicative of current or likely future behaviour. Using relatively long periods of recent behaviour, but applying fast temporal decay functions (so recent behaviour has a greater influence over the prediction than spatial behaviour in the more distant past) was found to be the most effective strategy.

In the example animation, a speed-heading prediction surface is generated for a journey by car from Cornwall to Hertfordshire (the driving scenario). The prediction calculates the likely future locations in one hour's time, based upon the spatial behaviour exhibited in the previous 3 hours. A moderately fast temporal decay function (power function with 1.5 exponent) is used to increase the influence of more recent points.

Temporal proximity predictions

Temporal proximity prediction surface animation
(0.16mb)

Predictions based upon temporal proximity build upon the principles of time geography and require a long-term history of previous spatial behaviour to be known. This approach encapsulates the region of space accessible to an individual, for a given time budget, based upon the locations that have previously been shown to be accessible from that location within some time limit. An approach which encapsulates previously visited locations using a buffer around paths was found to be more effective than a convex hull around all visited locations, which tended to enclose a greater proportion of unvisited, and hence possibly inaccessible, space. This approach was found to be most effective when predicting the location of individuals based upon their own long-term, previously exhibited behaviour in a region.

In the example animation, a temporal proximity prediction surface is generated for the daily commute to and from work for an individual living and working in zone 1, London (the daily migration scenario). The prediction calculates the likely future locations in 10 mins time, from analysis of the locations which have previously been accessible within 20 mins. A 250 metre buffer has been placed around the previously travelled space-time paths, to represent the accessible region.

Spatial proximity predictions

Spatial proximity prediction surface animation
(1mb)

Predictions based upon spatial proximity reduce the size of a prediction from the global to the local, but take no detailed account of the behaviour of the user, although varying the radius of the spatial proximity buffer, perhaps to reflect recently displayed speed, can dramatically improve results. The most effective surfaces tended to employ either slow distance decay functions, or no decay function at all, and perform best in situations where highly sinuous behaviour is observed (such as walking).

In the example animation, a spatial proximity prediction surface is generated for a walk up Helvellyn in the English Lake District; the walk describes a figure of eight (the walking scenario). The prediction calculates the likely future locations in one hour's time, using a spatial buffer of 3km with no distance decay.