Parallel Optimization And Application Research On Moving Object Detection And Recognition Algorithms

Computer vision and machine learning techniques have been wildly applied in many fields such as data mining,information security,remote sensing image processing,biology information,smart transportation,smart security and medical service and so on.As one of the most important computer vision technique,moving object detection has been broadly applied in many reality scene,existing moving object detection methods have long computing time and high complexity problems,however.How to satisfy the realtime requirement has become more and more important.With the rapid development of the deep learning technique,another computer vision technique,object recognition has achieved significant improvement on recognition accuracy performance compared to traditional methods.Nevertheless,due to the complex structure and high complexity,how to finish the training process rapidly has become an urgent problem in deep learning fields.With the advent of the ear of big data,a lot of data without label have been produced every day.Since the traditional supervised methods need label data to finish the training process and the cross-validation process for further application.However,it has become more and more difficult to add the label information by manual.It is of significance that unsupervised based clustering methods can fully utilize and explore the latent information.The performance of the clustering methods is critically dependent on the choice of affinity matric.Consequently,the research of fully utilize and explore the same input data for clustering is of critical importance in applications.The work in this thesis aims to address the above issue and its main contributions are summarized as follows.(1)We proposed a parallel optimization method in moving objects detection application.As for the moving objects detection application,we proposed a parallel optimization method based on the heterogeneous platform(CPU + GPU).This method fully utilize the parallel scheme in different platform and proposed the optimization strategy respectively.Moreover,a novel parallel pipeline implementation on a heterogeneous platform is also provided.(2)We proposed a parallel optimization method in object recognition algorithm,deep belief networks.According to the characteristic of deep belief network,we proposed a parallel optimization scheme based on the GPU hardware.This method fully utilizes the inherent two-level hierarchy of block and thread parallelism.Further,we present a new computation strategy for visible units and successfully solve the performance bottleneck problem.(3)We proposed a multi-kernel learning based multi-view clustering method.We study the spectral clustering under the multi-view setting and propose a robust multikernel learning based multi-view clustering algorithm.And we conduct comprehensive experiments to compare the proposed approach with existing state-of-the-art multi-view clustering methods on two benchmark data sets and the experimental results demonstrate the superiority of the proposed method.