Research Of Human Detection Based On Volterra Kernel Method

Human detection is an important research direction in the field of target detection identification. It is the basis of many computer vision applications such as pedestrian tracking, human gesture recognition, the crowd abnormal event detection, human behavior analysis, intelligent transportation, involving the fields of image processing, pattern recognition technology etc. So it’s very important to do the research of human detection to improve its accuracy and efficiency. On the basis of previous studies, this paper proposes a human detection method based on kernel methods, and improves human detection accuracy in two aspects.Based on the kernel method in classification problems, kernel function and the selection of the parameters have an important influence on classification results. But in the single kernel learning process, the selection of kernel function always based on experience, and kernel function parameter is determined by a lot of simulation experiments because there is no standard. In this paper, for the polynomial kernel function in GF space, we do certain change of its form under the norm restricted conditions, making the single kernel form into multiple’s, and utilize multiple kernel learning method to optimize weight coefficients. This method combines the simplicity of single kernel and strong adaptability of multiple kernels, so we don’t need to predefine parameters of kernel function. It can change kernel parameters according to the training samples, saving a large amount of simulation time, and improving the detection accuracy at the same time. In this paper, we use four optimization methods to optimize the polynomial kernel function in GF space, and the results show that under different ways of optimizing, the classification results are similar, so it confirms that the classification method we proposed has high adaptability. And the classification effect of the polynomial kernel function in GF space is better than popular used single kernel function and meets the multiple kernels’ match.In the process of human detection, we often have false negative due to people with similar shape and have false positive because of shade or complex background, which result in a not good detection effect. In order to solve these two problems, this article first make the human body divided into three parts:the upper part, left foot, right foot, and then add a full body to extract feature and train respectively, which can reduce the false negative rate. After complete the first stage of training, we carry out the second phase of training, which can reduce the false positive rate. And we train the multiple classifier in two stages, which can reduce the false positives rate. Then we use INRIA, TUD-Campus, and TUD-Crossing three dataset for testing. The experiment results show that using the classifier model in this paper can improve the detection rate, decrease the false negatives rate and false positives rate under the complex and crowed background.