Study On Target Tracking Method Based On SVM

In order to overcome the low precision shortcomings of the traditional target tracking system because of target's rotation, zoom, shelter and varied illumination etc, the support vector machine based on statistics learning theory is introduced into target tracking field in this paper. This alogrithm is able to detect and track automatically the pre-select target in watching field or image and gets over the most defects of traditional target tracking system.The algorithm theory and realization process of the proposed tracking method based on SVM are introduced in detail. Bary centre and correlation and those improved algorithms are implemented in this project in order to compare with the support vector machine (SVM) tracking method proposed in this paper. To resolve large computation and low tracking precision of taking raw image pixels value as input data of SVM training and classifing, the paper proposes three kinds of feature extract methods which provide input data for SVM. One of those methods is using Gabor wavelet to compute image feature and then using energy function, PCA and AdaBoost algorithm to extract small parts of Gabor features. The second method is using PCA and LDA to drop image feature dimensions. The last one is using Haar wavelet features and using AdaBoost to extract the representive part of Haar feature. It is proved in experiment that the Haar feature extracting based on SVM can achieve the best precision. Moreover, this proposed method is able to get realtime property through using cascade detect and track algorithm.According to the project requirement and the proposed algorithm property, a high speed target tracking hardware system based on single DSP is designed and used in actual project. The traditional matching method and improved methods are used in actual project and the proposed method has been tested. Due to the defect of large memory space and long run time of the proved SVM algorithm, this hardware system can't meet the requirement of real time. Thus, the author designs a dual DSP hardware system scheme which uses the currently highest performance DSP chipset...