Automatic Detection Of Traffic Lights Research

With the rapid economic development, the cars are becoming more and morepopular. However, when the cars bring convenience, they make the traffic conditionsbe complicated. More and more traffic lights are needed to provide navigationinformation at the corner of vehicles. The running intelligent vehicle must obtaintraffic light information to make its behavior decision. This dissertation presents aalgorithm of detection and classification for traffic lights. As is different from themainstream algorithm of detection, this system detects the backplane of the trafficlights first, and then recognizes the lights according to the relevant characteristics. Inorder to improve the speed, the algorithm crops the original image first, and deletesthe regions that have big difference with the backplane of the traffic lights on thevalue of color; and then detects with the correlation characteristic of the backplane toget the target regions. Finally, using the characteristics of the traffic lights to find itslocation and identify the type of the light-emitting element. In this dissertation, themain work and contributions are as follows.1. For the image after preprocessing, reasonable use of mathematical morphologyoperation of traffic lights back and connected to the rail, and according to the basicgeometric features of the back to filter the candidate regions.2. Using maximum likelihood ratio test based on principal component analysisalgorithm to distinguish the traffic lights back and impurities, to provide reasonableback the target area for subsequent light-emitting unit detection algorithm to locateclassification.3. When to distinguish the type of traffic light emitting unit, the Hu invariantmoment, Hough detection and the circular degree detection, this paper compares andanalyzes and select appropriate algorithm to extract the circular traffic lights; Forarrow traffic lights, the compared experiment measure method of template matchingand coordinate projection effect after choose the latter for rapid and efficientsegmentation.Collected pictures of real vehicle testing, the results verified the feasibility andefficiency in detecting the system at a red light. For image quality and other uncertain factors mistakenly identified targets, this dissertation is also given at theend of the next improvement ideas work.