An Integrated Traffic and Pedestrian Vision System

This page describes the Integrated Traffic and Pedestrian Vision System developed by a collaboration between the vision groups at Leeds and Reading Universities as part of the EPSRC's IMV program (EPSRC Grant GR/K46620).

The project combines two model-based tracking systems:

• Using active shape models to track non-rigid objects, in this case people. (Leeds)
• Using geometric 3D models to track rigid objects: in this case cars. (Reading)

Two complementary approaches for analysing the interactions between people and vehicles were explored:

• At Leeds we have shown how to detect atypical events using features of the spatio-temporal trajectories of people in relation to vehicles. To do this, the trajectories of people are compared with a probabilistic model for the distribution of these features obtained by observing large numbers of typical interactions.
• At Reading they have developed a system to produce natural language descriptions of object interactions using Bayesian belief nets.

Last Year I reviewed Daniel Crevier's book 'AI - the tumultuous history of the search for artificial intelligence' in which he relates how in 1966 Marvin Minsky set a first year undergraduate, Gerald Sussman, the task of connecting a television camera to a computer and getting the machine to describe what it saw. He did not succeed, but a joint project between Prof. David Hogg's group at Leeds University and Dr Tieniu Tan's at Reading (taking over the reins from the late Geoff Sullivan) comes close to achieving that goal. They have been able to automatically annotate image sequences to say, for instance, that 'Pedestrian n is now walking slowly in front of parked vehicle m' where n and m are simply the sequential numbers relating to appearance in the scene. Based on work David Hogg's group has done in finding algorithms to identify suspicious behaviour in car parks, it should not be long before the annotation can be extended to say 'Pedestrian n looks about to nick something out of vehicle m better have a human look at the image on the monitor'.

Fitting models to cars and people

Bigger
MPEG movie of tracking results (8M)

The Likelihood of a particular trajectory

postscript version
Jpeg version

MPEG movie (2MB) assessing the likelihood of a particular trajectories. Dots above the heads indicate likelihood of a particular trajectory. For this example the positions of cars were estimated by hand.

There are several other pages mainly meant for internal consumption, but you may find these of interest.

• Reading pages about the project
• Publications
• Leeds output data site
• Leeds tracking results
• VIEWS Reading work on which car tracker is based.

Web page created by Richard Morris rjm@comp.leeds.ac.uk 27 March 1998