基于YOLO v3的室内场景中人体检测方法的研究  

Studies on Human Detection Methods in Indoor Scenes Based on YOLO v3

作　　者：石强 葛源[1] 马树波 郑洪龙 韩娟娟 Shi Qiang;Ge Yuan;Ma Shubo;Zheng Honglong;Han Juanjuan(Nuclear Power Institute of China,Chengdu,610005,China)

机构地区：[1]中国核动力研究设计院,成都610005

出　　处：《仪器仪表用户》2021年第4期22-25,共4页Instrumentation

基　　金：贵州省科技计划项目(700283162107)。

摘　　要：针对传统人体检测算法,存在着鲁棒性差和对光照条件要求较为苛刻的问题,借鉴目标检测的最新研究成果,以YOLO v3网络为基础,对室内场景制做了相应的室内场景人体识别数据集;同时,结合人体成像具有宽高不一致的特点,聚类选取初始框的数量和规格,改进候选框在X轴和Y轴的分布密度,将红外夜视图片和常规图片混合训练,并利用运动检测算法提取ROI,然后进行噪声处理,最后进行了检测。实验证明,对室内场景的人体检测和追踪的方法比常用的人体识别方法具有更高的准确率、更低的漏检率。Aiming at the traditional human detection algorithm,there are problems of poor robustness and strict requirements on illumination conditions.Based on the latest research results of target detection,makes a corresponding family scene human body data set based on YOLO v3 network.At the same time,combined with the characteristics of wide and high inconsistency of human imaging,clustering selects the number and specifications of the initial frame,improves the distribution density of the candidate frame in the X-axis and the Y-axis,mixes the infrared night vision picture with the conventional picture,and uses motion detection.The algorithm extracts the ROI,then performs noise processing,and finally carries out the detection.The experiment proves that the method of human body detection and tracking for the home scene has higher accuracy and lower miss detection rate than the commonly used human body recognition method.

关 键 词：YOLO v3 室内场景 人体识别 聚类 混合训练 运动检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]