Research On Traffic Sign Detection And Classification Algorithms

Traffic regulation and safety proble m is getting more and more attention.Under this background, the concept of Intelligent Transportatio n System (ITS) ispresented. ITS integrates informatio n, communicatio n, control and comp utertechnologies into the existing transportation syste m, thereby establis hing a safe,efficie nt and reliable integrated transportatio n system. As a component of ITS,the traffic sign recognitio n system plays an important role in driver assistance,traffic s ign ma inta ining and automatica lly driving. However, in the complicatedtraffic scenes, the proble ms of different lighting condition, weather conditio n,partia l occlus ion, similar background color and shadow interfering make theresearch of tra ffic sign recognition far fro m mature. To address this s ituatio n,this thesis makes researches on hazy image restoration, traffic s ign detectio n andtraffic sign classification.Firstly, haze (inc luding fog, ha ze and dust) remova l algorithm isinvestigated in this thes is, and a new haze remova l algorithm based on sky regionsegmentation is proposed. As traffic sign recognition syste m usua lly works inoutdoor environment, it s hould be robust to various weather cond itions inc lud ingfog, haze and dust. By analyzing several existing algorithms, the interna lrelatio nship and shortcomings of the m are pointed out, and the haze remova lalgorithm based on sky region segmentation is proposed. The algorithm derivesfrom the dark channe l prior algorithm, and solves the origina l a lgorithm’sproblems of sky region distortio n, white object darken and ins ufficie nt ha loinhib ition. Experimenta l results show that the proposed algorithm can makeimages with fog, haze or dust significantly clearer, and thus improves theaccuracy of traffic sign detection and classification.Second ly, the proble m of traffic detection is thoroughly stud ied, and threefeatured traffic sign detection a lgorithms are proposed based on color, shape andmodel respectively. The task of tra ffic s ign detection is to find out a ll the regionsconta ining tra ffic s igns fro m the input ima ges. The proposed color basedalgorithm localizes prohibitory s igns by detecting red holes in the image, andeffectively solves the problems of similar background color and multiple signs clustered. The shape based algorithm localizes circular signs by detectingellipses in the ima ge. With the proposed ellipse detection algorithm based onsorted merging, circular signs can be detected fa st and accurately. The modelbased algorithm e mploys the multi-scale s liding window scheme. The detectionprocess cons ists of two steps na med the coarse filtering a nd the fine filtering, bywhich traffic signs with any color and any shape can be detected accurately.Experime ntal results show all the three algorithms give high detection accuraciesand are robust to adverse situatio ns such as bad lighting condition, occlusio n,and rotation, etc.Third ly, a coarse-to-fine traffic sign class ification algorithm is proposed.The task for traffic sign classification is to ana lyze the detected regions anddetermine the class of the sign in the region. By ana lys ing existing traffic signclassification algorithms, the major proble m a ffecting the classification accurac yis pointed out. Based on this ana lysis, a coarse-to-fine classification a lgorithm isproposed. The algorithm first classifies traffic signs into several super classes,then performs class-specific shape adjustment, and finally gets the fineclassification result. Experimenta l results show that the proposed algorithmoutperforms other existing a lgorithms in c lassification accuracy, and is robust tomany adverse situations.Lastly, for in-vehic le environments which have space and power limitation,the embedded imp leme ntation of tra ffic s ign recognition s ystem is studied, andtwo most time consuming modules are imple mented on FPGA (Fie ldProgrammab le Gate Array), which are the guided filtering module and the trafficsign detection module. The guided filtering module is used in hazy ima gerestoration. An effic ient, fle xible and low-latency VLSI (Very Large Sca leIntegration) architecture based on inte gral image is proposed, which requires nooff-chip me mory and can process300fra mes in one second for640×480images.The traffic sign detectio n module explo its the local property of circular Houghtransform, fully utilizes paralle l ability o f FPGA, and achieves prohibitory s igndetection with high speed and high accuracy. The module can process374fra mesof640×480images in one second.The proposed traffic sign detection algorithms have attended the GTSDB(German Traffic Sign Detection Benc hmark) competition held in2013, and ranked first in two categories and third in one category, refer to section3.6fordetailed results.The proposed traffic sign classification algorithm achieves a classificationaccuracy of99.52%on GTSRB (German Tra ffic Sign Recognition Benchmark),which is currently the best result for the dataset.