运动目标检测与特征提取算法的多层次并行优化

针对监控视频中运动目标实时特征提取的需求,在目标检测与特征提取串行算法的基础上,提出了基于OpenMP和多核CPU平台的三层并行优化算法。首先,在算法顶层,将串行算法抽象为两个模块组成的流水线,提出了流水线并行优化算法和相应的缓存管理策略;接着,在算法中层,考虑到特征提取模块中各子模块的功能独立性,设计了功能划分并行优化算法;最后,在算法底层,利用纹理特征提取模块的数据独立性,提出了数据划分并行优化算法。实验结果表明,该三层双模块并行优化算法在四核CPU平台上获得了接近Amdahl极限的加速比,基本实现了实际监控视频中运动目标检测与特征提取的实时处理。该多层次多模块并行优化方法普遍适用于串行算法在多核平台上进行并行优化的分析。

Multi-level Parallel Optimization of Moving Object Detection and Feature Extraction Algorithm

To meet the requirements of processing the surveillance moving objects in real time, a three-level parallel optimization algorithm of object detection and feature extraction based on multi-core platform and OpenMP is proposed. First, in the top level, the serial algorithm is separated into a 2-level pipeline, and a pipeline parallel algorithm and corresponding buffer management strategy are proposed; then, in the middle level, considering the functional independence of the sub modules of the feature extraction module, a functional decomposition algorithm is designed; lastly, in the bottom level, to utilize data independence of the texture extraction module, a data decomposition algorithm is proposed. Experiments show the proposed three-level and two-module parallel optimization algorithm acquires nearly the Amdahl speed up ratio, and realizes the real-time objective in surveillance videos. The multi-level parallel optimization method proposed in this article is common to be used in all parallel analyses of serial algorithms.