适用于密集人群目标检测的多尺度检测方法  

Multi-scale Detection Method for Dense Crowd Target Detection

作　　者：周建新[1] 吴建军[2] 薛均强 林帅[1] 党岗[1] 程志全 

机构地区：[1]国防科学技术大学计算机学院,湖南长沙410073 [2]中国人民解放军海南武警总队,海南海口570203 [3]湖南化身科技有限公司,湖南长沙410205

出　　处：《系统仿真学报》2016年第10期2503-2509,共7页Journal of System Simulation

摘　　要：针对密集人群场景下的目标检测问题,提出了一种多尺度的目标检测方法。在粗尺度下,使用优化的DPM(Deformable Part Model)检测方法,将人体全身作为检测对象,检测整个场景中的稀疏目标;在细尺度下,将头部作为检测对象,使用重新训练的Faster R-CNN(Region-based Convolutional Neural Network)网络检测稠密人群中的目标。将两种尺度下检测结果通过非极大值抑制(NMS,Non-Maximum Suppression)方法结合在一起,这样两种方法既互相补充又能去除冗余检测结果。实验结果证明,相比于单独的DPM检测方法和Faster R-CNN检测方法,提出的多尺度检测方法在检测精度上有显著提升。A multi-scale algorithm was proposed to detect the targets flexibly. In coarse scale, an optimized DPM （Deformable Part Model） method was used to filter out sparse objectives that was represented by whole body. Then the whole scenario was cut into multiple finer regions and the Faster R-CNN （Region-based Convolutional Neural Network） method was trained and utilized to detect dense objects that was indicated by head in fine scale. These two detection results were incorporated via NMS （Non - Maximum Suppression） method, in order to supplement with each other and remove redundancy. The effectiveness of the proposed method has been proved comparing detect accuracy with DPM and R-CNN individually in the final experiment.

关 键 词：密集人群检测 多尺度检测 DPM FASTER R-CNN 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]