作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (9): 240-251. doi: 10.19678/j.issn.1000-3428.0058800

• 图形图像处理 • 上一篇    下一篇

基于CenterNet目标检测算法的改进模型

石先让1, 苏洋1, 提艳1, 宋廷伦1,2, 戴振泳1   

  1. 1. 南京航空航天大学 能源与动力学院, 南京 210001;
    2. 奇瑞前瞻与预研技术中心, 安徽 芜湖 241006
  • 收稿日期:2020-07-01 修回日期:2020-08-24 发布日期:2021-09-13
  • 作者简介:石先让(1996-),男,硕士研究生,主研方向为自动驾驶、目标检测、模式识别;苏洋、提艳,博士研究生;宋廷伦,教授、博士;戴振泳,硕士研究生。
  • 基金资助:
    安徽省发改委重大研发项目“面向智能网联汽车的全线控底盘开发及测试验证)。

Improved Model Based on CenterNet Object Detection Algorithms

SHI Xianrang1, SU Yang1, TI Yan1, SONG Tinglun1,2, DAI Zhenyong1   

  1. 1. College of Energy and Power Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210001, China;
    2. Chery Advanced Engineering&Technology Center, Wuhu, Anhui 241006, China
  • Received:2020-07-01 Revised:2020-08-24 Published:2021-09-13

摘要: 在原CenterNet算法中,以Hourglass为Backbone的目标检测模型平均精度均值高于one-stage算法,但检测速度较低。为此,基于原有CenterNet目标检测算法,对Hourglass-104模型进行改进,设计一种Hourglass-208模型,并给出双特征金字塔网络特征图融合方法。在此基础上对目标大小和训练采用smooth L1损失函数,提出一种新的可端到端训练的目标检测算法T_CenterNet。在MS COCO数据集上的实验结果表明,该算法目标检测的评估指标AP50、APS、APM分别为63.6%、31.6%、45.8%,检测速度达到36 frame/s,综合性能优于原CenterNet算法。

关键词: 深度学习, 目标检测, Anchor-free方法, 关键点, 锚框

Abstract: In the original CenterNet algorithms, the object detection model with Hourglass as Backbone has a higher mean Average Precision(mAP) than other one-stage algorithms, but it is limited by the low detection speed.To address the problem, a new model named Hourglass-208 is proposed by using the original CenterNet object detection algorithm to improve the Hourglass-104 model.Additionally, a feature map fusion method for Twin Feature Pyramid Networks(TFPN) is given.On this basis, smooth L1 is used for the loss function of the object size to establish a new object detection algorithm, T_CenterNet, which can perform end-to-end training.Experimental results on the MS COCO data set show that the target detection evaluation index AP50, APS, APM of the proposed algorithm are 63.6%, 31.6%, 45.8%, respectively, and the detection speed of the algorithm reaches 36 frame/s.The comprehensive performance of the proposed algorithm is better than that of the original CenterNet algorithm.

Key words: deep learning, object detection, Anchor-free method, key points, anchor

中图分类号: