This project implements on an Android device a Traffic Sign Recognition(TSR) system capable of recognising 85 New Zealand traffic signs with variouscolours and shapes. I find it possible to have such system running almost realtime on medium-class Android devices by having the core processing modulewritten in native code, interacting with the user interface via JNI calls. Dif-ferent techniques are used to extract sign candidates from raw images beforethey are classified to the correct class. The first technique uses a colour filteroperating on either HSV or RGB colour space to filter out non-sign pixelsand the sign detector is applied on the remaining part of the images. Thesecond technique applies the detector on two consecutive raw images withoutcolour-filtering. Sign detector is a AdaBoost classifier using Linear BinaryPattern (LBP) to extract features. Products of sign detector are called ”signcandidates” and recognised by a back propagating neural network. This reportfinds the colour-filtering technique with the filtering threshold carefully tunedoverperforms the cascade detecting technique, but the latter is less likely tomiss a sign in different lighting conditions.

If you like this work, please consider sharing with , , , , or leave me a comment below.