Kernel Methods For Support Vector Machine In The Application Of Human Action Recognition

With the development of computer, computer vision technology has become a hotresearch topic in recent years. Computer vision is a kind of technology that usescameras or other photographic devices instead of people’s eyes to collect information,aiming to complete the process of moving target detection, tracking and recognitionthrough a computer instead of a human brain, so human action recognition becomesone of the hotspots in the field of computer vision. Human action recognition can beused in to make the human-computer interaction more intelligent, and it can also beused in intelligent video surveillance to make the video automatically understandinghuman behaviors come true.Support vector machine, proposed by Vapnik et al, is a new kind of machinelearning method based on statistical learning theory. It has obvious advantages interms of generalization capability and learning performance and can effectively solvethe problems of small sample, nonlinear and high dimension. Lots of experimentshave shown that support vector machine is an efficient method of classification.Support vector machine has been rapidly developed since the1990s and widely usedin many areas, such as face recognition, text categorization and medical imageprocessing.The kernel method is an important way to achieve nonlinear mapping in supportvector machine. The key technology of the kernel method is the selection orconstruction of the kernel function. This thesis focuses on the application ofmulti-kernel support vector machine in human action recognition.The main innovation of this paper is as follows:1. The local kernel function has a strong learning ability, but its generalizationperformance is weak. The global kernel function performs well in generalization, butlacks of learning ability. In order to overcome the weakness of the two kinds of kernelfunction and keep their advantages, we put forward a multi-kernel consisting of two kinds of kernel functions above. The experiments on human action recognition showthat multi-kernel function is better than single kernel function.2. The method of SVM based on multi-kernel function is proposed and put intothe experiments of human action recognition. Compared with template matchingmethod, this method has a small amount of computation and high rate of recognition.3. There are only color images in the database of human action from theUniversity of Illinois. In order to reduce the dimensions of images with the algorithmof principal component analysis and improve the speed of operation, all the colorimages are transformed into gray ones and establish a gray image database. Theresults of the experiments show that SVM based on multi-kernel function is practicaland effective in human action recognition.