Red means stop (sometimes)
Channa Meng, John M. Morris · 2017
Traffic lights (and most other traffic rules) are routinely ignored in Thailand ensuring Thai drivers kill themselves at the second highest rate of any country in the world. Hypothesizing that a red light runner (RLR) detection system, that was flexible enough to function effectively with poorly marked intersections and help police to educate RLRs, would save many lives, we designed a system that allows a single camera attached to a simple PC to monitor a controlled intersection and send images of vehicles violating traffic lights to a police post further down the road. Our system detects red lights and tracks vehicles, with sizes ranging from bicycles to trucks and buses, that `run' a red light. It first builds a Gaussian mixture model of the background that adapts to changing lighting conditions, then, in each frame, captures a mask of moving objects and segments it to identify and track individual vehicles. Templates derived from the masks are matched from frame to frame to track vehicles. Rules, based on predicted direction, size and shape, are used to resolve overlaps. We tested videos captured from a variety of intersections from over 7300 `red light' frames and correctly tracked more than 83% of events including vehicle merges and splits in them despite setting the camera only 1.5m from the road surface. All RLRs were effectively detected as a RLR only needed to be detected in one frame, with a camera positioned so that RLRs are foreground objects: even though the RLR sometimes merged with a background non-RLR, an effective image was obtained. Overhanging trees were sometimes not treated correctly as background using the Gaussian mixture model, so we added a post processing step, based on position and colour, to eliminate them from the lists of vehicles being tracked.