Recognition Of Chinese Traffic Lights Using On-board Vision Sensors

Recognizing Traffic Lights is one of the key tasks in realizing intelligent and more safe vehicles.On-board vision sensors are one of the most advantageous methods to recognize traffic lights with a high accuracy at low costs.The images taken by an onboard camera are processed by a traffic light recognition algorithm which detects and classifies potential traffic lights.Tsinghua University collected and labeled the first set of Chinese Traffic Lights which can be used to develop a traffic light recognition algorithm for Chinese Traffic situations.This algorithm consists of the parts image preprocessing,detection and classification,from which the classification stage is already developed.This research work focuses on implementing the detection algorithm based on feature extraction and a machine learning classification approach.A baseline detector is trained which serves as the basis for subsequent performance improvement steps.The developed detectors are compared based on their recall and false positives per sum of true positives and false negatives.Improvement approaches are based on increasing the number of positive training samples using image augmentation(flipping,rotating,scaling),increasing the number of training images by image augmentation(adding blur,increasing contrast,decreasing contrast),by modifying the number of weak classifiers,by adding padding and by modifying the learning tree depth of the classifier.This work concludes with an analysis of the compatibility of the developed detector with the existing classification stage.