Mobile Voronoi Diagrams for Traffic Monitoring under Bad Visibility Conditions

Abstract

A semiautomatic management of traffic scenes displays a large diversity of mobile data arising from usual Computer Vision techniques. The mobile nature of inputs requires the combination of different techniques for filtering, tracking, and clustering features along a video sequence. These problems are considerably harder in presence of low visibility conditions arising from rain, fog or dazzling conditions. It is necessary a robust coarse-to-fine approach for supporting early alert in presence of conflict or dangerous situation at road intersections. Currently, there is no a general solution developed for low visibility conditions, and what there is, has been developed following particular strategies involving a specific combination of filters for extracting and analyzing the situation. Under low visibility conditions, mobile features are clustered as blobs with similar motion patterns and labelled in terms of a mobile Voronoi site which represents the centroid of a coloured region with similar kinematic pattern. For a fixed camera, and in absence of information about relative velocities of vehicles, kinematic involves the relative variation of colour and shape. With low visibility conditions and for real-time response, it is not necessary to work with a large palette of colours, and a reduction of bits per pixel is performed in the preprocessing stage. We illustrate our results with some scenes where reflections in water (rainy weather) or discontinuities linked to fog, can produce hallucinations for which our approach provides a robust kinematic method justifying the application of mobile Voronoi diagrams for mobile blobs as unifying principle.