基于聚类与Faster RCNN的行人头部检测算法  

Pedestrian head detection algorithm based on clustering and Faster RCNN

作　　者：张洁 陈莉[1] 李铮 王森 陈昃 ZHANG Jie;CHEN Li;LI Zheng;WANG Sen;CHEN Ze(School of Information Science and Technology, Northwest University, Xi′an 710127, China)

机构地区：[1]西北大学信息科学与技术学院,陕西西安710127

出　　处：《西北大学学报（自然科学版）》2020年第6期971-978,共8页Journal of Northwest University（Natural Science Edition）

基　　金：陕西省重点研发计划项目(2019ZDLGY10-01,2019ZDLSF07-02);陕西省产业创新链项目(2017ZDCXL-GY-03-01-01)。

摘　　要：针对复杂环境中由于行人间相互遮挡导致检测准确率低的问题,考虑到行人头部与行人是一一对应关系,且头部在行人运动过程中不易被遮挡,提出了一种基于聚类与Faster RCNN的行人头部检测算法。设计一种新的距离度量方法,并结合k-means++算法对已标注人头检测框进行聚类,以确定anchor大小与长宽比;优化NMS算法惩罚函数剔除无效人头预测框,改善行人之间由于遮挡导致的召回率低的问题。仿真实验表明,该算法相比其他方法可有效提升行人头部检测精度,在Brainwash和SCUT-HEAD两个人头检测数据集上的最高AP值分别为90.2%和87.7%。Aiming at the problem that the pedestrian detection accuracy is low due to the pedestrian occlusion in a complex environment,considering that there is a one-to-one correspondence between the head and the person,and the head is not easily occluded during pedestrian movement,a pedestrian head detection algorithm based on clustering and Faster RCNN is proposed.The k-means++algorithm used the newly designed distance measurement method to cluster all labeled head detection boxes and determine the anchor size and aspect ratio.The penalty function of the NMS algorithm is optimized to remove invalid head prediction boxes,which can alleviate the problem of low recall due to pedestrian occlusion.The experiments show that compared with other methods,the proposed algorithm effectively improves the detection accuracy of pedestrian head.The highest AP on Brainwash and SCUT-HEAD datasets reached 90.2%and 87.7%respectively.

关 键 词：行人检测 聚类 Faster RCNN 非极大值抑制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]