计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (1): 194-199.DOI: 10.3778/j.issn.1002-8331.2005-0343

• 模式识别与人工智能 • 上一篇    下一篇

改进YOLOv3的非机动车检测与识别方法

叶佳林,苏子毅,马浩炎,袁夏,赵春霞   

  1. 南京理工大学 计算机科学与工程学院,南京 210094
  • 出版日期:2021-01-01 发布日期:2020-12-31

Improved YOLOv3  Non-motor Vehicles Detection and Recognition Method

YE Jialin, SU Ziyi, MA Haoyan, YUAN Xia, ZHAO Chunxia   

  1. School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
  • Online:2021-01-01 Published:2020-12-31

摘要:

随着交管部门对非机动车监管力度的增强,在道路交通监控视频中检测和识别非机动车将逐渐成为电子交警系统的必备功能。由于非机动车密度大,容易互相遮挡,且在监控视频中所占面积往往较小,容易出现检测定位不准确和漏检等问题。针对非机动车检测定位不准确和漏检问题,基于YOLOv3,提出一种改进的非机动车检测与识别模型,通过设计新的特征融合结构降低非机动车漏检率,使用GIOU损失提高定位准确度。实验结果表明,所提出的改进模型在自建真实复杂场景非机动车数据集上取得了优于YOLOv3的检测结果,将检测的平均检测准确率(mAP)提高了3.6%。

关键词: 非机动车检测, YOLOv3, 特征融合, GIOU损失

Abstract:

With the strengthening of the supervision of non-motor vehicles by the traffic management department, the detection and identification of non-motor vehicles in the road traffic monitoring video will gradually become an essential function of the electronic traffic police system. Due to the high density of non-motor vehicles, it is easy to block each other, and the area occupied is often small in the surveillance video, which is prone to problems such as inaccurate detection and missed detection. Aiming at the problems of inaccuracy and omission of non-motor vehicle detection, based on YOLOv3(You Only Look Once), an improved detection and recognition model for non-motor vehicles is proposed. By designing a new feature fusion structure, the detection rate of non-motor vehicles can be reduced, and the location accuracy can be improved by using GIOU(Generalized IOU) loss. The improved model obtains better detection results than YOLOv3 on the real complex scene non-motor vehicle data set, and the mean Average Precision(mAP) of detection is improved by 3.6%.

Key words: non-motor vehicle detection, You Only Look Once(YOLOv3), feature fusion, Generalized IOU(GIOU) loss