基于改进的Cascade-RCNN网络的人员检测算法  

Person detection algorithm based on improved Cascade-RCNN network

作　　者：吉鹏飞 JI Pengfei(Information Institute Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息学院,杭州310018

出　　处：《智能计算机与应用》2021年第8期107-111,117,共6页Intelligent Computer and Applications

基　　金：国家自然科学基金(6207050141);浙江省自然科学基金(LQ20F050010)。

摘　　要：为解决人员密集情景下行人检测存在大量的目标误检、漏检的情况,本文提出了一种改进的基于Cascade-RCNN的目标检测网络,提高了人员检测的准确率。对目前检测效果较好的Cascade-RCNN做了一些改进:选用ResNeXt101代替ResNet作为骨干网络,以便提取更加充分的特征;为了获得更好的标记框,用kmeans聚类算法得到更符合目标形态的anchor长宽比例,通过WBF算法融合多个模型的结果得到更精确的边界框,同时引入多尺度训练以提高对小尺度目标的检测能力。实验结果表明,在CrowdHuman公开数据集上,用ResNeXt101提取特征其得分提高了3.7%,用kmeans聚类算法生成anchor比例和WBF算法融合多预测框其准确率提升了0.7%和1.2%,最终整体性能较基础Cascade-RCNN提升近6%。In order to solve the problem of a large number of false and missed target detections in pedestrian detection in crowded scenarios,an improved target detection network based on Cascade-RCNN is proposed to improve the accuracy of human detection.Some improvements have been made to Cascade-RCNN,which has good detection results:ResNeXt101 is used instead of ResNet as the backbone network to extract more sufficient features;in order to obtain better labeled frames,the kmeans clustering algorithm is used to obtain anchor lengths that are more in line with the target shape Wide scale,the results of multiple models are merged through the WBF algorithm to obtain a more accurate bounding box,and multi-scale training is introduced to improve the detection ability of small-scale targets.Experimental results show that on the CrowdHuman public data set,using ResNeXt101 to extract features increases the score by 3.7%,using kmeans clustering algorithm to generate anchor ratios and WBF algorithm fusion multi-prediction frame,the accuracy rate increases by 0.7%and 1.2%,and finally the overall The performance is improved by nearly 6%compared to the basic Cascade-RCNN.

关 键 词：行人检测 Cascade-RCNN kmeans聚类算法 WBF算法 多尺度训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]