Research For Object Recognition Algorithm Under Urban Complicated Circumstance

Object detection is one of the hot topics in the field of digital image processing and pattern recognition, and it has the extensive area of application. Support vector machine (SVM), based on the statistical learning theory, keeps to the structural risk minimization principle, and effectively solves the over fitting problems while other traditional pattern recognition methods could not, and thus becomes the preferred classifier in the field of pattern recognition. This dissertation whose research objects are vehicles and foot passengers under the complicated traffic scenarios of urban roads is based on the intelligent vehicle active safety forewarning system. In this dissertation, we researched the application of Support vector machine in the field of object recognition and presented a new method of object recognition which was based on the support vector machine and ensemble learning.First, we detected the object such as vehicles and foot passengers under urban traffic circumstance. After we transformed the color image to gray image and did some noise removal, the vertical gradient components of image were found by using the vertical mask of Sobel operator. Then these candidate areas with sufficient amount of edges were determined by calculating the gradient of edges. As symmetry was an important feature of object, the proposed method measured the prominent vertical symmetry of candidate areas and verified all possible candidates after finding possible vehicle candidates. This method also figured out the vertical axis of symmetry.Combining the characteristic of edge, shadow and proportion between height and width, the contour rectangle of the object would be detected near the axis of symmetry. Then, AdaBoost-SVM multi-class classification algorithm based-on the mixture of kernel function was presented and applied to the urban object recognition. To obtain a set of weak learners which have the appropriate accuracy and diversity, this algorithm used mixture of kernels, established by the combination between polynomial kernel function and RBF kernel function, to be the kernel-function of SVM, and combined AdaBoost to make adaptive modulation for kernel-parameters, and then this set of weak learners were weightedly combined to gain a strong learner. Before making recognition to the detected objects, some samples of urban object were acquired, and their characteristics that include invariance, texture and proportion between height and width were gained to compose character vectors. After that, these vectors using the multi-class classification algorithm which is presented in this dissertation were trained and the strong learner was obtained. The detected objects would be the new unknown samples and eventually were tested to ascertain the sort they were affiliated with.The result of experiment indicates that the algorithm can satisfy the practical requirement because of the high accuracy and efficiency of recognition.