计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (8): 20-25.DOI: 10.3778/j.issn.1002-8331.1907-0153

• 热点与综述 • 上一篇    下一篇

改进YOLOv3的全景交通监控目标检测

孔方方,宋蓓蓓   

  1. 长安大学 信息工程学院,西安 710064
  • 出版日期:2020-04-15 发布日期:2020-04-14

Improved YOLOv3 Panoramic Traffic Monitoring Target Detection

KONG Fangfang, SONG Beibei   

  1. School of Information Engineering, Chang’an University, Xi’an 710064, China
  • Online:2020-04-15 Published:2020-04-14

摘要:

针对城市交通场景复杂、车辆及行人等目标多且尺度变化大等特点,提出一种改进的YOLOv3全景交通监控多目标检测方法。以YOLOv3网络为基础,兼顾大小尺度目标特性设计4个检测尺度,并进行多尺度特征融合处理。利用[K]-means聚类方法对数据集中的标注目标框进行聚类分析,选取优化的聚类锚点框宽高维度作为改进YOLOv3网络的初始候选框。全景交通监控检测目标包括大型汽车、小型汽车、骑行摩托车、骑行自行车和行人5类。在测试集上目标检测平均精度和召回率分别达到84.49%和97.18%,较原始YOLOv3分别提高了7.76%和4.89%,处理速度可满足交通场景下实时性检测要求。

关键词: 目标检测, 全景交通监控, [K]-means, 特征融合, 卷积神经网络

Abstract:

Aiming at the complex traffic scenes, vehicles and pedestrians, and the large scale changes, an improved YOLOv3 panoramic traffic monitoring multi-target detection method is proposed. Firstly, based on the YOLOv3 network, four detection scales are designed with consideration of the size and scale target characteristics, and multi-scale feature fusion processing is performed. Then [K]-means clustering method is used to cluster the labeling target box in the dataset, and the width and height of optimized clustering anchor box are selected as the initial candidate box. The panoramic traffic monitoring targets include five categories:large vehicle, small vehicle, motorcycles, bicycles and pedestrians. The average accuracy and recall rate of target detection on the test set reaches 84.49% and 97.18%, respectively, which is 7.76% and 4.89% higher than the original YOLOv3. The processing speed can meet the real-time detection requirements in traffic scenarios.

Key words: target detection, panoramic traffic monitoring, [K]-means, feature fusion, convolutional neural network