| Cascade detector is one kind of strong classifiers, which is composed of several weak classifiers in a certain order. Cascade detectors are attractive for their simplicity and efficiency and have been widely used in the field of fast target detection. This thesis focuses on constructing the cascade detector based on linear SVMs and the main contributions are concluded as follows:1. Cascade structures are analyzed in detail. One method of constructing a cascade classifier with traditional linear SVMs is proposed. The learning problem of every node in the cascade structure is described as a new quadratic programming problem in the framework of SVMs, which makes every linear classifier achieve very high detection rate but only moderate false positive rate. The experiments show that this method has good generalization capacity and much fast speed compared with the traditional SVMs.2. The precondition of using the new cascade classifier constructed by our method is discussed. Furthermore, some measures are presented to deal with some cases which do not meet the precondition.3. To speed up the training process, a new algorithm is presented to construct a cascade of linear classifiers with L-SVMs (Lagrangian Support Vector Machine, L-SVM): First, the negative data are divided into several parts according to the geometric distribution of the training data. Here, every parts of negative data can be separated from the positive data with a linear classifier; Second, a linear L-SVM is used to obtain the linear classifiers between every part of negative data and positive data; Lastly, the linear classifiers are combined to construct a cascade detector. Experiments show that the new algorithm is much faster than the cascade classifier based on traditional SVMs in the training process, and the classification accuracy is almost the same.4. The cascade classifiers are applied to grass recognition and pedestrian detection, and are proved to be efficient. |
PDF Full Text Request