Study On The Algorithm Of Digital Marks Recognition For CyberCar

CyberCar (a regional intelligent vehicle) is a low-speed electric and smart unmanned mini-vehicle. It can transport persons and goods in crowded regions, raise the utilization rate of resources,alleviate traffic congestion,and reduce noise and environmental pollution to a certain extent. (CyberCar has been put into a trial operation phase abroad). Therefore,we developed a vision-based navigation CyberCar. LED digital encoding stations are set up,but tests showed that there are some difficulties about CyberCar's digit recognition ability to adapt to environment changes and anti-deformation based on gray images under complicated outdoor environment. So this paper studies digit segmentation and recognition based on color LED digital coding. Its main content are as follows:1. Designed a digit recognition program based on color information. At first, the area of interest is identified based on the navigation path interval and priori knowledge. Secondly, we can verify the existence of digit through hue component in HSI color space. Then the illumination condition is figured out through the illumination classifier which was designed in past. Finally, digit recognition is realized using image segmentation and pattern recognition.2. Studied color images segmentation algorithms. As for color digits,this paper presents two color digital image segmentation algorithms: region growing and clustering algorithms. Seeds selection,growth rules and ending conditions were researched based on region growing principles. Seeds selection: The seeds H and S values were derived through statistical extreme values of H and S component in a certain area with a priori knowledge. Pixels whose values are equal to the seeds'H and S values were regarded as seeds in the process of scanning a image; Growth rules determined by average value vectors and the variance vectors of neighbor pixels of seeds; when all seeds growth meet the conditions of regiongrowing ending rules,image segmentation is completed. This method is based on regional consistency of color vector utilizing spatial information as well. According to clustering algorithms,the initial clustering center selection,similarity criteria and clustering criteria are researched. The initial clustering center can be accessed by calculating the extreme values of the two-dimensional histogram of H and S component in a certain area; Similarity criteria is determined by the distance of two color vectors in HS plane; Clustering algorithms end when clustering centers do not change. This method guarantees that same cluster data has a higher similarity,different cluster data has a lower similarity. On this basis,we analyze and compare these two approaches and find that region growing method has a better result.Meanwhile, different image preprocesses are utilized according to different illumination conditions to enhance the adaptive ability of image segmentation algorithms in this paper. In normal illumination conditions , digital image segmentation is completed by using region growing or clustering algorithms directly; In strong illumination conditions,Algorithm of Histogram Cone-shaped Extending is performed on H and S component of images firstly as H component and S component of images have lower contrast; In weak illumination conditions,there are some noises in the image. Vector intermediate value filter method and Algorithm of Histogram Cone-shaped Extending are performed for the same reason. Then digital image segmentation is completed by region growing or clustering algorithms.3. Studied LED digit recognition methods of decision tree based on morphology and template matching based on digital structure. Erosion algorithm is used to withdraw the structure characteristics of digits,then digit recognition is realized by using the decision tree. Through analyzing LED digits,a conclusion is drawn that digits can be recognized by digits'structure characteristics. Then two morphology templates are constructed which are used to detect vertical and horizontal structure characteristics of digits. Decision tree is established using two kinds of structure characteristics, so digits are recognized. As for the second recognition algorithm: Firstly,thinner algorithm is used to obtain the skeleton of the object. Secondly,crossing points and extreme points are detected through defined templates and the number of circles is obtained by region growing. Digits are roughly classified into four feature spaces based on those characteristics. Digitsspace one include extreme points only,digits space two include circles only,digits space three include extreme points and cross points,and digits space four include extreme points and circles. In these spaces,digits are finally recognized by using template matching method based on other characteristics.4. Verified segmentation and recognition algorithms. Software programs are developed using Visual C++ to realize the algorithms'functions. Experiments results of color digits recognizing under different illuminations prove that algorithms of this paper is reliable and robust.