Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (9): 162-167.DOI: 10.3778/j.issn.1002-8331.2002-0126

Previous Articles     Next Articles

Research on Deep Neural Network in Multi-scene Vehicle Attribute Recognition

WANG Lin, CHAI Jiangyun   

  1. School of Automation and Information Engineering, Xi’an University of Technology, Xi’an 710048, China
  • Online:2021-05-01 Published:2021-04-29

深度神经网络在多场景车辆属性识别中的研究

王林,柴江云   

  1. 西安理工大学 自动化与信息工程学院,西安 710048

Abstract:

Single vehicle attribute recognition cannot meet the existing traffic system. In order to improve the vehicle detection and positioning reliability in actual monitoring, a model is built using the idea of deep neural networks, which can identify vehicle attributes in two different scenarios:close range monitoring scenarios and traffic monitoring scenarios, included vehicle type and color. Based on the YOLOv3 neural network, it is improved to reduce the network depth while ensuring the accuracy rate. The vehicle type and color attributes are graded training, improve model detection speed. In addition, the AttributesCars vehicle attribute dataset is created to complete the data preparation. Experimental results show that the proposed method can meet the real-time requirements of video under the premise of an average accuracy rate of 95.63%, and has achieved good results in two different scenarios, is suitable for multi-scenario vehicle attribute recognition.

Key words: deep neural network, close range monitoring scene, traffic monitoring scene, YOLOv3, vehicle type, vehicle color

摘要:

单一的车辆属性识别已无法满足现有的交通系统,为了提高在实际监控中车辆检测定位的可靠性,利用深度神经网络的思想建立了一种能够在近景监控场景和交通监控场景两种不同场景下识别车辆属性的模型,主要包括车辆类型和颜色两种属性类别。以YOLOv3神经网络为基础,对其进行改进,降低网络深度的同时保证准确率,将车辆类型和颜色属性进行分级训练,提高模型检测速度,此外,创建了AttributesCars车辆属性数据集完成数据准备工作。实验结果表明,所提方法在平均准确率为95.63%的前提下可以满足视频的实时性要求,并且在两种不同场景下均取得了不错的成绩,适用于多场景车辆属性识别。

关键词: 深度神经网络, 近景监控场景, 交通监控场景, YOLOv3, 车辆类型, 车辆颜色