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 ›› 2020, Vol. 46 ›› Issue (6): 26-33. doi: 10.19678/j.issn.1000-3428.0056462

• 热点与综述 • 上一篇    下一篇

基于改进卷积神经网络的交通场景小目标检测

郑秋梅, 王璐璐, 王风华   

  1. 中国石油大学(华东) 计算机科学与技术学院, 山东 青岛 266580
  • 收稿日期:2019-10-31 修回日期:2019-12-03 发布日期:2019-12-04
  • 作者简介:郑秋梅(1964-),女,教授,主研方向为模式识别、图像处理;王璐璐(通信作者),硕士研究生;王风华,讲师、博士。
  • 基金资助:
    国家自然科学基金(61305008);国家自然科学基金面上项目(51274232);中央高校基本科研业务费专项资金(19CX02030A);山东省自然科学基金面上项目(ZR2018MEE004)。

Small Object Detection in Traffic Scene Based on Improved Convolutional Neural Network

ZHENG Qiumei, WANG Lulu, WANG Fenghua   

  1. College of Computer Science and Technology, China University of Petroleum(East China), Qingdao, Shandong 266580, China
  • Received:2019-10-31 Revised:2019-12-03 Published:2019-12-04

摘要: 针对复杂交通场景中的小尺度车辆检测问题,提出改进的YOLOv3目标检测方法(S-YOLOv3)。使用ResNet网络优化YOLOv3的Darknet-53特征提取结构,采用特征金字塔网络获取目标的4个尺度特征以融合浅层特征和深层特征信息,并根据检测目标的大小调整损失函数的影响权重,从而增强小目标及相互遮挡物体的检测效果。在KITTI数据集上的实验结果表明,S-YOLOv3方法的检测速度和平均精度均值分别为52.45 frame/s和93.30%,相比YOLOv3方法在保证小目标检测实时性的同时具有更高的检测精度。

关键词: 改进的YOLOv3方法, 特征提取, 多尺度融合, 损失函数, 小目标检测

Abstract: In view of the problems existing in the detection of small vehicles in complex traffic scenes,this paper proposes an improved objection detection method based on YOLOv3,S-YOLOv3.The method uses ResNet to improve the feature extraction structure of Darknet-53 of YOLOv3.The four scales of features of the object are obtained by using the Feature Pyramid Network(FPN) to fuse shallow feature information and deep feature information.Then the impact weight of the loss function is adjusted according to the size of the detection target,so as to enhance the detection performance of small objects and occluded objects.Experimental results on KITTI dataset show that the average detection speed of S-YOLOv3 is increased to 52.45 frame/s and Mean Average Precision(mPA) is increased to 93.30%.Compared with the YOLOv3 method,this proposed method can improve the precision of small object detection while ensuring the real-time performance.

Key words: improved YOLOv3 method, feature extraction, multiscale fusion, loss function, small object detection

中图分类号: