基于改进YOLOv2模型的驾驶辅助系统实时行人检测  

Real-time Pedestrian Detection in Advanced Driver Assistance Systems Based on Improved YOLOv2 Model

作　　者：白中浩[1] 李智强 蒋彬辉[1] 王鹏辉 Bai Zhonghao;Li Zhiqiang;Jiang Binhui;Wang Penghui(Hunan University,State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body,Changsha 410082)

机构地区：[1]湖南大学汽车车身先进设计制造国家重点实验室

出　　处：《汽车工程》2019年第12期1416-1423,共8页Automotive Engineering

基　　金：国家自然科学基金(51621004,51475153);福建省汽车电子与电驱动重点实验室开放基金(KF-X18001)资助

摘　　要：为解决驾驶辅助系统(ADAS)对复杂背景行人和小尺寸行人检测精度较低的问题,基于深度神经网络模型YOLOv2建立了ADAS实时行人检测模型YOLOv2-P。首先在特征提取网络中采用参数化修正线性单元激活函数,以从训练数据中自适应地学习参数,并在行人检测网络中采用多层特征图融合方法,将低层特征图信息与高层特征图信息进行融合;然后使用交叉熵损失函数替代YOLOv2模型中的sigmoid激活函数,并对宽度、高度损失函数进行归一化处理;最后采用迭代自组织数据分析算法对行人数据集中行人边界框尺寸进行聚类。试验结果表明:相比于YOLOv2,YOLOv2-P对复杂背景行人及小尺寸行人的检测精度有明显提升,能够满足ADAS行人检测准确性和实时性需要。In order to solve the problem of low detection accuracy of pedestrians with small stature or in complex background in advanced driver assistance systems(ADAS),a real-time pedestrian detection model named YOLOv2-P for ADAS is established based on a deep neural network model YOLOv2.Firstly,the parametric rectified linear unit activation function in feature extraction network is adopted to adaptively learn parameters from training data,and the multi-feature map fusion method is used to fusion low-layer feature map and high-layer feature map in feature extraction network.Next,cross entropy loss function is used instead of sigmoid activation function in model YOLOv2,and the width and height loss functions are normalized.Finally,a clustering is performed on the pedestrian boundary frame size in pedestrian data set by utilizing iterative self-organizing data analysis algorithm.The results of test show that the detection accuracy of the pedestrians with small stature or in complex background with YOLOv2-P model has a significant rise,compare with that with YOLOv2 model,meeting the requirements in accuracy and real-time performance for the pedestrian detection in ADAS.

关 键 词：行人检测 驾驶辅助系统 参数化修正线性单元 交叉熵损失函数 迭代自组织数据分析算法 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]