Adaboost算法的并行化及其在目标分类中的应用

发布时间：2018-01-30 22:28

本文关键词： GPGPU MIC 目标分类自适应增强学习车辆识别　出处：《华南理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：目标分类处于智能视频监控分析系统中的关键环节,因为当我们需要对视频序列中的运动目标进行监测时,要先检测、再分类,最后才是运动轨迹分析和理解。其中,Adaboost(Adaptive Boosting,自适应增强学习)分类算法应用最为广泛,其核心思想是针对同一个训练集训练不同的弱分类器,然后把这些弱分类器集合起来,构成一个更强的最终分类器。但是要想获得性能较好的Adaboost分类器,往往需要花费大量的时间在样本训练上。而且训练算法需要占用较大的内存空间,普通家用电脑由于自身计算能力方面能力的限制,Adaboost算法在普通电脑上难以开展。为了降低传统Adaboost算法训练时间,本文结合当下主流的多核协处理器——MIC和GPGPU,针对Adaboost特点开展了相关的并行优化工作:1、对传统Adaboost算法进行热点分析,发现Adaboost算法在训练弱分类器时,90%以上的耗时集中在特征值计算及排序上面;2、结合协处理器不同的硬件架构及编程风格,在GPGPU平台上采用并行双调排序方式优化排序,改变了原有数据的存储方式,减少随机访存的时间,提高了样本训练速度。此外,本文还针对MIC平台开展了相应的并行优化实验,对热度相对集中的函数使用MPI/OpenMP等并行编程工具进行并行化,且细分为粗粒度并行以及细粒度并行两种优化策略。在样本数为25600,样本大小为18*18,MIC与GPGPU分别获得3.8和7.2加速比。实验表明,GPGPU在处理图像数据方面更为出色。为了提高算法分类准确率,在样本选取方面,本文提出了新型的样本采集方式,在对车辆进行识别时,极大提高了原算法的检测精度。此外,针对目标分类检测过程中速度较慢的问题,提出了并行优化方案。
[Abstract]:Target classification is a key step in the intelligent video surveillance and analysis system, because when we need to monitor the moving targets in video sequences, we must first detect and then classify. Finally, the analysis and understanding of motion trajectory. Among them, Adaboostan Adaptive boost (Adaptive Enhancement Learning) classification algorithm is the most widely used. Its core idea is to train different weak classifiers for the same training set, and then set these weak classifiers together. To form a stronger final classifier, but to obtain a better performance Adaboost classifier. It often takes a lot of time to train samples, and the training algorithm needs to occupy a large amount of memory space. The common home computer is limited by its own computing ability. In order to reduce the training time of traditional Adaboost algorithm, this paper combines the current mainstream multi-core coprocessor, MIC and GPGPU. According to the characteristics of Adaboost, the parallel optimization work: 1 is carried out. The focus of traditional Adaboost algorithm is analyzed, and it is found that Adaboost algorithm is training weak classifier. The time consuming over 90% is concentrated on the calculation and ranking of eigenvalues; 2. Combined with the different hardware architecture and programming style of the coprocessor, the parallel double-tone sorting method is adopted to optimize the sorting on the GPGPU platform, which changes the storage mode of the original data and reduces the time of random memory access. The training speed of the sample is improved. In addition, the parallel optimization experiments are carried out for the MIC platform. The functions with relatively concentrated heat are parallelized by parallel programming tools such as MPI/OpenMP and subdivided into coarse-grained parallelism and fine-grained parallelism. The number of samples is 25600. The sample size is 18 ~ (18) mics and GPGPU is 3. 8 and 7. 2 speedup ratio, respectively. The experiment shows that. In order to improve the classification accuracy of the algorithm, in the aspect of sample selection, this paper proposes a new way to collect samples, which is used to identify vehicles. The detection accuracy of the original algorithm is greatly improved. In addition, a parallel optimization scheme is proposed to solve the problem of slow speed in the process of target classification detection.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN948.6;TP301.6

【参考文献】