针对畸变行人检测的神经网络  

A Multi-Class Pedestrian Detection Network for Distorted Pedestrians

作　　者：张姣 肖江剑[2] 周传宏[1] Zhang Jiao;Xiao Jiangjian;Zhou Chuanhong

机构地区：[1]上海大学机电工程与自动化学院,上海200000 [2]中国科学院宁波工业技术研究院,浙江宁波315000

出　　处：《计量与测试技术》2018年第9期40-44,共5页Metrology & Measurement Technique

摘　　要：针对视角存在畸变的视场下行人会依据所在视场中的位置表现不同于常规的多种姿态,如何在资源有限的情况下高效高精度的进行行人检测提出一种"多类"目标检测算法,即将视场中不同畸变程度下的行人可视为不同类型的监测目标,并依托当下分类算法中表现优异的Faster R-CNN神经网络框架,在安防监控图像中大大提高了行人检测的精度和速度。算法由训练分类器和检测两部分组成。在训练阶段,添加对数据集做多分类的分类层,即将不同畸变程度的行人定义为不同类的检测对象,很大程度上避免了训练的时候忽略因畸变造成的共性特征,以此提高检测精度。其次,以梯度下降的速度即收敛速度作为我们是否找到精准的多分类"边界"的依据,换句话说,我们定义的多分类的边界是不断变化的。在检测阶段,我们对不同场景实践展示出最终的检测结果。结果证明:不论是速度,精度都证明了多分类思想的正确性。In case of the field with perspective distortion,pedestrians show different attitudes from postures as usual. How to get high accuracy of pedestrian detection under limited resources,we propose a ＂ multi class＂ target detection algorithm. We regard different pedestrian under field with different visual distortion as different targets.Based on faster R-CNN neural network,we improve the speed and accuracy of pedestrian detection. The paper is consisted of two parts,training and testing. In the first stage of training,the pedestrian in data set is defined as different types by the distortion. In this way,we avoid to ignore the common character caused by the distortion. Secondly,the gradient descent velocity is basis whether we find ＂ precise boundary＂,in other words,the boundary is changing. In the detection part,we practice in different scenes and show the final detection results. The results show that： whatever the speed and precision are proved and ＂ multi class＂ is meaningful.

关 键 词：行人检测 视角畸变 多分类 神经网络 多目标检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]