• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于集成学习与位置信息约束的前方车辆检测

耿磊1,2,彭晓帅1,2,肖志涛1,2,李秀艳1,2,甘鹏1,2   

  1. (1.天津市光电检测技术与系统重点实验室,天津 300387;2.天津工业大学电子与信息工程学院,天津 300387)
  • 收稿日期:2017-07-16 修回日期:2017-11-09 出版日期:2018-10-25 发布日期:2018-10-25
  • 基金资助:

    国家自然科学基金(61601325,61771340);天津市自然科学基金(17JCQNJC01400)

Preceding vehicles detection based on integrated
learning and constraint of positions information

GENG Lei1,2,PENG Xiaoshuai1,2,XIAO Zhitao1,2,LI Xiuyan1,2,GAN Peng1,2   

  1. (1.Tianjin Key Laboratory of Optoelectronic Detection Technology and Systems,Tianjin 300387;
    2.School of Electronics and Information Engineering,Tianjin Polytechnic University,Tianjin 300387,China)
  • Received:2017-07-16 Revised:2017-11-09 Online:2018-10-25 Published:2018-10-25

摘要:

针对传统前方车辆检测方法难以同时满足准确性与实时性问题,提出一种结合AdaBoost集成学习与位置信息约束的车辆检测方法。首先,利用Edge Boxes算法根据车辆边缘序列信息计算推荐窗口。然后,通过帧存坐标系中车辆位置信息对非目标推荐窗口进行排除。最后,将过滤后窗口聚类处理并择优选取作为AdaBoost分类器输入,进行检测评判,并对最终检测结果进行边框回归处理,以提升检测精准度。实验结果表明,该方法对于不同检测场景有较强鲁棒性,能够同时满足车辆检测的准确性与实时性要求。
 

关键词: 前方车辆检测, 集成学习, 位置信息, 边框回归

Abstract:

Since the traditional preceding vehicle detection strategies cannot meet the accuracy and real-time requirements simultaneously, we propose an approach which combines AdaBoost, an integrated learning method, with the constraint of positions information. Firstly, regions proposal (RP) are obtained by edge boxes method according to the sequence information of vehicle edges. Secondly, the position information of vehicles in the  frame coordinate system is used to filter out nontarget RPs. Finally, the obtained windows are clustered and fed into the AdaBoost classifiers for vehicles detection, and at the same time borders regression is utilized to improve the accuracy of detection results. Experimental results demonstrate that the proposed method has robustness to different detection scenarios and that it can meet the accuracy and real-time requirements of vehicle detection.
 

Key words: preceding vehicles detection, integrated learning, position information, borders regression