Computer Engineering and Applications ›› 2019, Vol. 55 ›› Issue (7): 138-144.DOI: 10.3778/j.issn.1002-8331.1807-0209

Previous Articles     Next Articles

Pedestrian Detection Algorithm Based on Motion Feature and Position Estimation

GONG Jianfeng1, HAN Jiandong1, DENG Yifang2   

  1. 1.School of Computer and Information Technology, Shanxi University, Taiyuan 030006, China
    2.Software Engineering Division, Shanxi Institute of Automation, Taiyuan 030012, China
  • Online:2019-04-01 Published:2019-04-15



  1. 1.山西大学 计算机与信息技术学院,太原 030006
    2.山西省自动化研究所 软件工程事业部,太原 030012

Abstract: According to the characteristics of the movement of pedestrians and the correspondence between position and height of pedestrians in images, the pedestrian detection algorithm combining motion feature with location estimation is proposed. The motion feature and the Aggregated Channel Features(ACF) are extracted, then these features are trained by the classifier, and an evaluation model is built on the possible position of pedestrians. In the detection stage, the candidate pedestrian areas are determined by the classifier, and then the non maximum suppression algorithm is used to eliminate redundant windows. Finally, the model of position evaluation is used to judge the pedestrian candidate regions to eliminate non-target pedestrian.  In the experiment, the Caltech data set is used in the simulation of the algorithm, the results show that our algorithm achieves a Log-Average Miss Rate(LAMR) of 27.12%. Compared with the ACF algorithm, LAMR is decreased by 5.9%. The results of the experiment testify that motion feature can be used as a complementation of static feature, and location estimation of pedestrian has certain ability to distinguish when the camera is fixed.

Key words: motion feature, location estimation, ACF feature, Real Adaboost

摘要: 根据行人运动的特点和行人在图像中位置与身高的对应关系,提出了一种结合运动特征与位置估计的行人检测算法。提取运动特征和聚合通道特征(ACF),将提取的特征放到Real Adaboost分类器里进行训练,并对行人可能存在的位置建立评估模型;在检测阶段,通过分类器确定行人的候选区域,然后采用非极大值抑制算法去除重叠窗口,最后对行人候选区域应用位置评估模型进一步判断,以此排除可能的非行人目标。实验采用Caltech数据集对算法进行仿真,该算法的平均对数漏检率为27.12%,与ACF算法的相比降低了5.9个百分点。实验表明运动特征在视频检测中能够与静态特征进行信息互补,同时行人的位置估计在摄像机固定的情况下,具有一定的判断能力。

关键词: 运动特征, 位置估计, ACF特征, Real Adaboost