The Research For Detection And Recognition Of Moving Vehicle In Intelligent Transportation

In the domain of Intelligent Transportation System, no matter it is Transportation Management System, Integrate Transportation Hinge Coordinate System, Leading Information Service System, Vehicle Managing and Intelligent Control System, the detection and classification recognition of moving vehicle is one of basic tasks. Traditional moving vehicles detection methods, such as electromagnetism interaction, laser, infrared and ultra sound, suffer from a number of drawbacks and unable to satisfy the accurate requirements. In recent years, with the developments of computer and digital technique, video image detections get hold of attentions and are widely used in the detection of moving vehicles due to simple, steady and exact. In this thesis, the research for detection and classification recognition of moving vehicle is processing from video image detections and tracks new field of intelligent system. The detection and classification recognition of moving vehicle is a difficult and arduous task for integrated application of computer technology, pattern recognition, image processing, applied mathematics and theory of vision. Based on analysis and summarized of related and down to date research results in-country and out-country, this thesis aims at detection and classification recognition of moving vehicle on the condition of complex background that contains other moving vehicles and shadow of vehicle and accomplishes a series of analysis of arithmetic, research of theory and actual application. Main content of this thesis includes two kinds: moving vehicle detection and moving vehicle classification recognition. To detect moving vehicle from complex background and overcome the influences of other moving vehicle and vehicle shadow, the moving vehicle segmentation method based on deformable template model is proposed. For that can't build or do not build beforehand, a HHM segmentation method based statistical layered model is put forward as a complementary. In classification recognition of vehicle, to reflect structural and applied characteristic of every kinds of vehicle from characteristic data of vehicle, a clustering method based CEF and information potential energy is used to classify vehicle. On the base of classify, rough set and neural network method is used to enable vehicle recognition system have the capability in study and recognition. In one word, the main aim of this thesis integrated image processing, computer vision, pattern recognition, numerical value processing, statistical analysis, rough theory, neural network, and combined with computer application, and to provide some basal and valuable methods and algorithms for the realization of intelligence system. The main context of this thesis can be summarized as follows: (1) The detection of moving vehicle on complex background In practicality, there are many vehicles moving on the way and these vehicles may be blocked off each other and produce some moving shadow. It can't eliminate moving background and moving shadow using one image subtracting another image. So, it difficult to segment and extract interested vehicle from images. In this case, this thesis's primary contribution lies in: We designed a moving vehicle segmentation method based ondeformable template model Vehicles can be expressed by some given shape parameter models because vehicles have definite geometry shape features. Different kinds of vehicle present the change of deformable parameters. Having finished the study and training process of geometry shape features of many vehicle samples, this thesis established several representative deformable template models and parameter constraint functions. Using the parameters of models deforming and the optical algorithm, when the contours of deformed template model encompasses only those pixels that moving vehicle and the optical algorithm attains maximum value, the moving vehicle segmentation is accomplished. That is called match procession. In the optical algorithm, for decreasing the computing time, a traditional EM method is amended and a geometrical anneal temperature schedule is put forward. Compared to the traditional logarithmic anneal temperature schedule, the robot and computing time are improved. Through the analysis of the mechanism and experimental results of this segmentation method, we can get the conclusion that this method can extract the interested vehicle from the complex image background that contains moving objects and shadow. It need explain that, except for segmentation, this method can also be used to classify vehicle. So it is a very simple and efficient method of segmenting and recognizing of moving vehicle. We proposed a HHM segmentation method based statistical layered model For an image including interest vehicle, other moving object and shadow of moving objects, this thesis proposed statistical layered model. In statistical layered model, the interest vehicle is called foreground layer and other moving object is called background layer and shadow of moving objects is called shadow layer and is expressed by thestatistical model respectively. The model parameters are estimated by the HMM-based method of video sequences. This method need not study beforehand and can study straightway in the process of segmentation. So it can overcome the shortcoming of deformable template model method. HMM-based method makes use of the spatial relativity and time relativity of video sequences to accomplish recognition of model. The experimental results show that this method can succeed segmenting the moving vehicle. (2) The classification and recognition of vehicle The classification and recognition of vehicle is a popular problem. Following the Chinese standard, vehicle is classified in big(8-10ton),middle(2-8ton)and small(<2ton)or truck(van), passenger car(car) and special vehicle. But it is so simple and almost does not be used in practice. So in this thesis we classify vehicles by the habit of people: a truck, a van, a two compartments car, a three compartments car, a mini-car, a business-car, a Santana, a Buik, a Mazda, a Ford, a Toyota and so on. Thus the main research of vehicle classification and recognition in this thesis are as follows: Proposed a clustering method based CEF and information potential energy is used to classify vehicle Clustering is one of the important methods in pattern recognition. Information entropy is an obvious criteria to establish the clustering rule in vehicle classification because it can dictate the structure of the data. This thesis proposed a clustering algorithm using clustering evaluation function of entropy-based. When applied to nonlinearly character data of vehicles, the algorithm performs the classification of vehicles well. Compared to KNN and other methods, the method's result is better. Proposed a based rough set and neural network vehicle study andrecognition method Rough set is a new means to be used in the classification of blur, imprecise and having incompletion data information. It's primary idea is to maintain the classification ability unchanged by the way of knowledge reduction and to bring up the decision-making and classification rule. So it can predigest the dimension of recognized object and decrease the computing time. Neural network is a good method to study in patter recognition. Combining RS with neural network does vehicle classification. The results shows that the methods has the advantages of fast computation and easy realization, and the method is better than other methods in optimized attributes, decrease the indexes of analysis and test samples to a great extents, and enhancing the right ratio of classification. Detection and recognition of moving vehicle is a basis step in ITS. It has too difficulties because the shape of vehicles are multiform and constantly change. But we belief that IT's future is more beautify.