上海大学学报(自然科学版) ›› 2019, Vol. 25 ›› Issue (1): 56-65.doi: 10.12066/j.issn.1007-2861.1889

• 研究论文 • 上一篇    下一篇

一种基于单目视觉的前方车辆检测算法

陈高攀1, 徐美华2(), 王琪1, 郭爱英2   

  1. 1. 上海大学 微电子研究与开发中心, 上海 200444
    2. 上海大学 机电工程与自动化学院, 上海 200444
  • 收稿日期:2017-02-24 出版日期:2019-02-28 发布日期:2019-02-26
  • 通讯作者: 徐美华 E-mail:mhxu@shu.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(61376028);上海经济和信息化委员会资助项目(11XI-15)

Front-vehicle detection method based on monocular vision

CHEN Gaopan1, XU Meihua2(), WANG Qi1, GUO Aiying2   

  1. 1. Microelectronic Research and Development Center, Shanghai University, Shanghai 200444, China
    2. School of Mechatronic Engineering and Automation, Shanghai University,Shanghai 200444, China
  • Received:2017-02-24 Online:2019-02-28 Published:2019-02-26
  • Contact: XU Meihua E-mail:mhxu@shu.edu.cn

摘要:

针对前方车辆检测中遇到的速度和精度问题, 首先对感兴趣区域 (region-of-interest, ROI) 提取中大津阈值 Qtsu 分割方法应用范围的局限性进行了分析, 并考虑累计密度函数 (cumulative-density-function, CDF) 阈值分割方法鲁棒性差的问题, 提出了一种基于人工神经网络 (artificial neural network, ANN) 的方法来拟合阈值分割点, 以提高阈值分割的准确性和鲁棒性, 减少伪 ROI; 其次针对在复杂环境下车辆与背景粘连的问题提出了一种分区域六角分割的方法, 降低 ROI 漏检率; 最后对 ROI 提取 Haar 特征以及用优化的自适应提升 (adaptive boosting, AdaBoost) 对 ROI 进行分类, 得到筛选后的车辆目标. 实验结果表明, 对于分辨率为 640$\times $480 的视频, 系统处理速度可达到 81.3 帧/s, 错误率为 3.8%.

关键词: 车辆检测, 感兴趣区域, 人工神经网络, Haar, AdaBoost

Abstract:

To speed and improve accuracy of front-vehicle detection, limitations of Qtsu threshold and poor robustness of cumulative-density-function (CDF) threshold in the extraction of region-of-interest (ROI) are analyzed. An artificial neural network (ANN) method is used to fit threshold point to improve accuracy and robustness of the threshold and reduce false ROIs. A region-hexagon-segmentation method is presented to deal with overlap between vehicle region and background in a complicated environment for reducing misdetection. Haar features in the ROIs are extracted, and the ROIs are classified with an optimized adaptive boosting (AdaBoost) classifier to get real vehicle target. Experiments show that processing speed using the proposed method is 81.3 frames per second for a 640$\times $480 video with an error rate of 3.8%.

Key words: vehicle detection, region-of-interest (ROI), artificial neural network (ANN), Haar, adaptive boosting (AdaBoost)

中图分类号: