Traffic Surveillance: Sistema Avançado de Vigilância de Tráfego Rodoviário
Project Type: General Research Project
Research Field: Mobile Robot Sensing
Sponsors: FCT – MCT – PRAXIS/EEI/11218/1998
Time span: 1999-2002

The authors propose to build an automatic traffic surveillance system which will learn by observation the rules drivers use to control their vehicles. The system will be able to assess a driver's performance relative to that of his/her peers and to diagnose the reasons for any discrepancy. The system will consist of a number of stationary surveillance cameras connected to ordinary personal computers. The computers will track any vehicles in and feed the results of the tracking into an established machine learning program. By assuming that any acceleration of a car (i.e. speeding up, slowing down or turning) indicates a control input to the car the learning program will be able to associate the driver's actions with the car's current situation (i.e, the state of any nearby vehicles, etc.).

The proposed project will produce a car tracker with capabilities significantly in advance of those presently available. Cars will be tracked with multiple cameras and over large areas with a tracker able to detect the cars' signal lights The emphasis has been on increasing the reliability, robustness and range of applicability of the tracker. Real-time performance of the trackers would be a welcome capability, and one to aim at eventually, but the system goals can all be met by slower off-line processing.

The results of the tracking will be compared to independently measured data in order to validate the tracker. An inertial sensor for robot navigation has been developed and will be available to this project, and an agreement in principle from the Coimbra traffic police to use their doppler-radar equipped surveillance cameras as been achieved.

The proposed project will deliver a software system able to track cars over large areas using multiple communicating trackers and learn the rules used by the drivers of those vehicles. The cars turn and brake signals will be recovered and tracking will be possible in a greater range of weather conditions than possible at present. It is anticipated that the models learnt by the system will be usable to assess and diagnose the performance of either individual or groups of drivers.

Related People

Jorge Dias
Jorge Lobo