The Design Of Illumination Classifier For Lanes

Safety is always one of the most important topics in vehicle engineering. In recent years with the rapid development of transportation, traffic accidents of highway, especially the malignant traffic accident, are greatly increasing. Under this circumstance, Safety Driving Assist technologies, as a part of Intelligent Vehicle technologies, are paid more and more attention. Safety Driving Assist System uses the information from Intelligent Apperceiving System and makes decision and programming. Then the system gives advice to the driver or replaces the driver to control the vehicle partially.Lane Departure Warning System is an important part of Safety Driving Assist System. At present Lane Departure Warning System are used as electrical fittings and amounted on the vehicles in some countries such as America and Japan. But in China, there are not Lane Departure Warning Systems which have independent intellectual property. JLUVA-LDWS System was exploited by Intelligent Vehicle Group of Jilin University. It studies lane mark recognition and tracking under different illumination such as night, weak light, common light and strong light. Experiment results show that the algorithms have good performances.As an important part of Lane Departure Warning System, the lane illumination classifier proposed in this paper can adaptively classify an image into a corresponding sort, then transfer corresponding algorithms to recognize lane marks.As a pattern recognition system, lane illumination classifier mainly consists of two parts: the extraction of the features and classification recognition. So the paper includes four parts: 1. the extraction of the features; 2. the design ofBack-Propagation Neural Network; 3. the design of fuzzy classifier; 4. the design of Support Vector Machines.Choosing and extraction good features have great influences to the performance of a classifier. This paper uses two methods: the method based on experiences and the method based on K-L transformation. The former algorithm is easy and has good real-time ability. But it has subjectivity because the extraction of the features is depended on the designer's experience. The latter is a common method in the field of pattern recognition; it can eliminate the relevancy among the features. The method based on K-L transformation can reduce the feature space to a low dimension in the precondition of certain variance percentage. But the calculation is complicated and it suffers from the problems of large memory requirement and CPU time.The second part of the paper is the design of BP Neural Network Classifier. The basic principle and related knowledge are introduced first; then use the features which were obtained in the Chapter 2 as the input of the network. Thus build network. After comparison, Levenberg-Marquardt algorithm is chosen as the training method of the network. Select perfect network model by adjusting the structure parameters especially the number of hidden nerves cell time after time. The network model can converge rapidly and uses little iterative time because it adopt improved training algorithm. Besides, the results of 40 samples show that the network has good Generalization Capability. Finally, we input the network weight and threshold into the Visual C++ environment, thus we can realize the BP neutral network classifier in the Visual C++ environment.Because the illumination is a continuous concept, So we proposed the classifier that based on fuzzy sets theory. In this paper, we found a fuzzy classifier by using the fuzzy toolbox. Select three features as the input language variable, select triangle as the basic figure of the membership functions. Then determine the detailed figure and location of the membership functions. Finally determine four fuzzy inference principles,and realize the fuzzy classifier.The statistical learning theory and Support Vector Machines are discussed in the last part of the paper. The statistical learning theory is a machine learning principle that aimed at small scale samples, while SVM algorithm is an implement mode of the Statistic Theory. It has good sample learning precision and the learning space is complicated. So it is the implement embodiment of the structural risk minimization principle ERM. This paper founds Support Vector Machines. Use Gauss kernel function as inner product function, use one-to-one strategy to solve the multi-classification. Finally obtain 46 support vectors (120 samples) in 6 classifiers by adjust error punishing parameter C perpetually, thus constitutes decision classification cover. 40 samples validate the advantages of SVM in the aspect of small scale samples.In conclusion, this paper completed the design of Back-Propagation Neural Network classifier,fuzzy classifier,SVM classifier. Use the same testing samples to validate the algorithms. Finally choose the classifier that based on BP Neural Network and realize the BP classifier Visual C++ environment. Thus the purpose of this paper is achieved.