We propose a robust segmentation system for road traffic signs using adaptive thresholds. The main objective is to develop a robust segmentation system under various illumination conditions. The main problem is how to obtain the best suited threshold for relative similarity. It can be got automatically by K-nearest connectivity when K=4 and the two-way connectivity method. Our system utilizes relative similarity of neighborhood pixels and shape information of traffic signs together. The effectiveness of our proposed system is confirmed through experiments under various illumination conditions. In the experiments, 90 images taken under daytime, evening and night-time are used. The system can give 98% successful segmentation rate.