Parallel Design And Implementation For Intelligent Video Analysis On GPU

Intelligent video analysis technology is developing rapidly in recent years as a new generation of computer vision technology. However, the computational complexity of the intelligent video analysis algorithm and the data density of video data make it difficult to process video in real time. General-purpose computing on GPU technology has wide-ranging practical application value, and has achieved remarkable acceleration effect in many parallel computing areas. Therefore, the technology of intelligent video analysis based on GPU parallel computing has important research meaning.In this thesis, the parallel design and implementation for intelligent video analysis on GPU is studied. The innovative work includes:By analyzing the processing flow of intelligent video analysis, the intelligent video analysis function is divided into high or low level according to the depth of information mining. For the moving target detection and tracking technology which is a low level intelligent analysis technology, the GPU parallelization scheme for moving target detection based on background subtraction method is designed and implemented, a GPU based parallel moving target tracking scheme based on mean shift is also designed. The schemes accelerate the processing speed of the initial critical steps in the intelligent video analysis, and provide the basis for the real-time application of the function on higher level.As a high level intelligent video analysis technology, the pedestrian re-identification is divided into three subproblems, which are pedestrian detection, feature extraction and feature matching. A GPU parallel framework for pedestrian re-identification is proposed by studying the GPU parallel technology of pedestrian detection, feature extraction and feature matching. The GPU parallel schemes for pedestrian detection based on HOG+SVM is proposed, which is optimized on GPU. A feature extraction technique of the assemblage feature for pedestrian recognition is proposed by introducing color self-similarity, HOG and LBP features, and the GPU parallel extraction technique is studied in detail. After designing the principle of similarity matching, the similarity computation and the feature matching algorithm are implemented on GPU. The experiments verify the effectiveness and real-time performance of the framework.