基于GPU的视频流人群实时计数

发布时间：2018-04-11 18:31

本文选题：视频监控 + GPU并行计算　；参考：《计算机应用》2017年01期

【摘要】：为了解决人群遮挡严重、光照突变等恶劣环境下人群计数准确率低的问题,提出基于混合高斯模型(GMM)和尺度不变特征变换(SIFT)特征的人群数量统计分析新方法。首先,基于GMM提取运动人群,并采用灰度共生矩阵(GLCM)和形态学方法去除背景中移动的小物体和较密集的噪声等非人群前景,针对GMM算法提出了一种效率较高的并行模型;接着,检测运动人群的SIFT特征点作为人群统计的基础,基于二值图像的特征提取大大减少了执行时间;最后,提出基于人群特征数和人群数量进行统计分析的新方法,选择不同等级的人群数量的数据集分别进行训练,统计得出平均单个特征点数,并对不同密度的行人进行计数实验。算法采用基于GPU多流处理器进行加速,并针对所提算法在统一计算设备架构(CUDA)流上任务的有效调度的方法进行分析。实验结果显示,相比单流提速31.5%,相比CPU提速71.8%。
[Abstract]:In order to solve the problem of low accuracy of population counting under severe occlusion and sudden change of illumination, a new method of population quantitative analysis based on mixed Gao Si model (GMMM) and scale-invariant feature transform (sift) was proposed.Firstly, based on GMM extraction, gray level co-occurrence matrix (GLCM) and morphological methods are used to remove the non-crowd foreground of moving small objects and dense noise in the background. A parallel model is proposed for the GMM algorithm.The detection of SIFT feature points of moving population is the basis of population statistics, and the feature extraction based on binary image greatly reduces the execution time. Finally, a new statistical analysis method based on the number of population features and the number of people is proposed.The data sets of the number of people of different grades were selected for training, and the average number of individual feature points was obtained by statistics, and the counting experiment of pedestrians with different density was carried out.The algorithm is accelerated based on GPU multi-stream processor, and the efficient scheduling method of the proposed algorithm on the unified computing device architecture is analyzed.The experimental results show that compared with single flow and CPU, the speed is increased by 31.5 and 71.8, respectively.
【作者单位】：上海理工大学光电信息与计算机工程学院;上海理工大学管理学院;
【基金】：国家自然科学基金资助项目(61572325,60970012) 高等学校博士学科点专项科研博导基金资助项目(20113120110008) 上海重点科技攻关项目(14511107902) 上海市工程中心建设项目(GCZX14014) 上海市一流学科建设项目(XTKX2012) 沪江基金研究基地专项(C14001)~~
【分类号】：TN948.6;TP391.41

【相似文献】