Researches And Applications Of Multiple Kernel Support Vector Machine And Classifier Ensemble Method In Intelligent Transportation Systems

In the recent development of urban construction, Intelligent Transportation Systems(ITS) are playing more and more important roles. With the development of traffc data collection technology and the increment of the hardware capacity for data storage, there is massive traffc data stored in the urban traffc management center. How to find the useful knowledge from the massive data is an urgent problem to be solved for the traffc managers. Data mining is a new research area which mines useful information from the massive data. This thesis is devoted to study how to apply the data mining technologies, including multiple kernel support vector machine and classifier ensemble method, and computer vision technologies to improve the function and operation effciency of ITS. It focuses on three key problems of ITS:(1) traffc incident detection;(2) traffc flow speed estimation and traffc index computation and visualization of dynamic road network areas;(3)vehicle logos recognition. The major contributions of this thesis are as follows:1. We have investigated multiple kernel support vector machine(MKL-SVM) and its application in automatic incident detection(AID). Support Vector Machine(SVM) has wide applications in many fields. For ITS, it has distinctive features, which are:(1) it generates a huge amount of data which includes many noise;(2) the distributions of the numbers of the incident and nonincident sample are very unbalanced. Therefore, the classical SVM method can not ensure AID algorithms get good stability and high accuracy. Besides that, the performance of SVM depends on the kernel function and its parameters very much. If the kernel and its parameters are selected appropriately, SVM will have good performance; On the country, the performance is bad. Until now, researchers have to try many times for obtaining the appropriate kernel and its parameters. This is a trivial way, and there is not a structured way to choose them. In order to solve these problems, we propose a method, MKL-SVM,to detect the traffc incidents. This method adopts the approach that summing the general kernels with weights to avoid selection of the appropriate kernel and its parameters, and itemploys gradient descent method and SMO(Sequential minimal optimization) algorithm to solve MKL-SVM effciently. Experimental results show that MKL-SVM has robust performance and it can avoid the burden of choosing the appropriate kernel and its parameters.More importantly, MKL-SVM has better average performance comparing to SVM.2. In order to improve the performance of MKL-SVM further, we propose multiple kernel support vector machine ensemble(MKL-SVM ensemble) method to detect traffc incidents.Meanwhile, we compare MKL-SVM ensemble to SVM, SVM ensemble and MKL-SVM method. Experimental results show that MKL-SVM ensemble has better average performance than MKL-SVM, and it has the best comprehensive performance and the performance is very robust.3. We have researched multiple kernel support vector regression(MKL-SVR) and its applications on traffc flow speed estimation and traffc index computation and visualization of dynamic road network areas. Traffc flow state identification is an important function of ITS. Especially, the congested state identification is an important guarantee of traffc flow guidance and the smooth running of urban traffc. Speed is taken as a key input for ITS to identify the state of traffc flow. In Shanghai, most of the industrial loop detectors(ILDs)are installed in a single loop way, and these ILDs can only detect the parameters of flow,saturation, etc.. However, the speed can not be detected directly. If the relationship between flow and speed can be mined accurately, then we can obtain the speed by the flow. For this purpose, this thesis investigates the traffc speed estimation methods which include polynomial fitting algorithm, BP neural networks, and Support Vector Regress(SVR). In order to improve the speed estimation accuracy, we propose a new method, MKL-SVR, to estimate the speed. Experimental results show that MKL-SVR has high estimation accuracy and it is very robust. Then, based on the MKL-SVR speed estimation results, we compute and visualize the traffc indexes for dynamic road network areas. We have completed the goal of computing and visualizing the traffc indexes in our laboratory intelligent transportation system, which can a?ord useful traffc information for the travelers and traffc managers.4. We propose a new global image feature extraction algorithm using the sharpness histograms.In order to extract the features from vehicle logo images e?ectively, this thesis firstly proposes an adaptive algorithm for corner detection based on the degree of sharpness of the contour. This method only uses the geometry information of the edge structure to detect corners. It calculates the sharpness degree threshold firstly, then takes the points whose sharpness degree lager than the threshold as the corner candidates. Then it propose a method of making a projection for an edge segment toward the edge's fitting line to remove salientpoints. Finally, if some neighboring corner candidates are detected in a support area, it o?ers an e?ective scheme to merge them into one point, and the final corners are reserved. Based on this corner detection method, this thesis propose a novel global image feature extraction algorithm using the sharpness histograms, which extracts the edges form the image firstly,then computes the sharpness degree. At last, it obtains the sharpness histograms.5. We have investigated the method of weighted multi-class support vector machine ensemble and its application in vehicle logos recognition. This thesis presents a new method to recognize the vehicle logos by weighted multi-class support vector machine ensembles and compares this method to K-Nearest Neighbor and multi-class SVM. Experimental results indicate that our method has improved the vehicle logo recognition accuracy greatly and has very robust performance comparing to the other two methods. Unlike multi-class SVM method, our method can avoid the burden of choose the appropriate kernel and its parameters.