计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (22): 233-238.DOI: 10.3778/j.issn.1002-8331.1804-0265

• 工程与应用 • 上一篇    下一篇

基于级联微型神经网络的多角度车辆检测方法

李  浩1,2,连  捷3,王辛岩4   

  1. 1.西安文理学院 西安市物联网应用工程重点实验室,西安 710065
    2.东南大学 交通学院,南京 210000
    3.中国电子科技集团公司 第三十八研究所,合肥 230088
    4.西藏大学 工学院,拉萨 850000
  • 出版日期:2018-11-15 发布日期:2018-11-13

Multi-view vehicle detection method based on cascade micro neural networks

LI Hao1,2, LIAN Jie3, WANG Xinyan4   

  1. 1.Laboratory of IOT Key Technologies and Applications, Xi’an University, Xi’an 710065, China
    2.School of Transportation, Southeast University, Nanjing 210000, China
    3.No.38 Research Institute, China Electronics Technology Group Corporation, Hefei 230088, China
    4.School of Engineering, Tibet University, Lhasa 850000, China
  • Online:2018-11-15 Published:2018-11-13

摘要: 车辆检测是智能交通系统建设的关键步骤,但在光照变化、遮挡等复杂交通场景下,单一角度视频检测的方法无法准确地获取车辆特定特征。为了提高交通监控图像中车辆检测的准确性,将AdaBoost算法嵌入微型的神经网络模型,并结合局部归一化像素差值特征(LNPD),提出了基于级联微型神经网络的多角度车辆检测方法。该方法首先提取检测图像的局部归一化像素差值特征,然后使用多层感知器学习最优的特征子集及其组合特征,最后使用AdaBoost算法筛选最具区分力的特征构建强分类器。以不同复杂程度的真实交通场景中包含有正面、侧面及背面三个角度的样本集作为测试集,并与NPD、DPM-V5、ACF和RCNN等方法进行了定性与定量对比。实验结果表明,该车辆检测方法在三种数据集上的平均检测率和检测时间分别为82.28%和125 ms,优于传统车辆检测方法。

关键词: 智能交通, 车辆检测, 微型神经网, LNPD特征, 级联分类器

Abstract: The vehicle detection is a critical step for intelligent traffic system construction, while the detection method using single view cannot accurately detect the vehicle specific features in the complex traffic scene such as illumination change and shelter. To improve the accuracy of vehicle detection in traffic surveillance images, embedding the AdaBoost algorithm in the micro neural network and combining Local Normalized Pixel-value Difference(LNPD) features, this paper proposes a multi-view vehicle detection method based on cascade micro neural network. The method first extracts local normalized pixel-value difference features for images and then learns optimal feature subset and its combination features using multilayer perceptron. Finally, the AdaBoost algorithm is employed to filter discriminative features and construct a strong classifier. In the three real traffic scenarios with different complexity, a dataset with front, side and back three viewpoints is used as a test set, and the test results are compared with NPD, DPM-V5, ACF and RCNN adopting qualitative and quantitative analyses. The experimental results show that the proposed method is optimal with an average detection accuracy of 82.28% and an elapsed time of 125 ms.

Key words: intelligent transportation, vehicle detection, micro neural networks, LNPD feature, cascade classifier