Research On Traffic Light Recognition For Intelligent Vehicle

As the current research focus of automotive industry, intelligent vehicles improve ride comfort and safety through the auxiliary or substitute of driver actions. But driverless technology is not reliable enough, and it takes a long term to update the current transport facilities. So at this stage, intelligent vehicles mainly carry driver assistance systems to achieve the intelligent functions. As a part of the driver assistance system, traffic lights recognition based on machine vision technology is essential to guarantee the safety o f intelligent cars driving on city roads. This paper studies on recognition of traffic signal lights and vehicle brake lights both in daytime and nighttime. Methods to identify different conditions are designed and validated with Matlab.Firstly, this paper designs a method for the recognition of circular traffic signal lights in daytime. The original image is filtered with a threshold value of RGB color space, with the result of which, a filtering method based on morphological characteristics of each area is done. The recognition of arrow traffic signal lights is designed. With the threshold value of RGB color space and local gray feature,the complete arrow-shaped region is extracted. With the extracted candidate region, arrow-shaped traffic lights are recognized by morphological characteristics and center shift ratio of each region.Secondly, a method is designed for the special conditions for nighttime recognition. The original image is filtered with the threshold values of RGB color space. The screening of the number of zero pixels surrounded by regions and local area negated operation are conducted to the color filtered image to eliminate the irregular halo surrounding the lights. With the filtering method based on morphological characteristics of each area, and round shaped region searching method based on Canny edge detection and Hough transform, circular traffic signal lights are recognized.Lastly, methods are designed for vehicle brake lights recognition for daytime and nighttime. For daytime recognition, original image is filtered with both threshold values of RGB color space and gray feature, and the candidate regions are enhanced by morphological dilation. For nighttime recognition, the original image is filtered with threshold values of YCb Cr color space. Light candidate regions are extracted with Top-Hat transform combined with Otsu’s method. With the results of color filter and light candidate region extraction, morphological operations and filters are conducted to extract the brake light regions. Recognition is completed with a pairing process to mark brake light regions belong to the same vehicle, for results of daytime and nighttime.Experiments are conducted for recognition of traffic signal lights and brake lights. Results of experiment indicates the methods designed could recognize traffic signal lights and brake lights effectively, but wrong results appear under some complicated condition. Some improvements are still needed.