基于DeR-FCN模型的车辆检测算法  

Vehicle detection algorithm using DeR-FCN model

作　　者：王玲 李厚博 王鹏[1] 孙爽滋[1] WANG Ling;LI Hou-bo;WANG Peng;SUN Shuang-zi(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机工程与设计》2020年第10期2927-2933,共7页Computer Engineering and Design

基　　金：吉林省科技发展计划技术攻关基金项目(20190302118GX)。

摘　　要：针对复杂城市环境下天气、光照、目标尺度以及车辆之间的遮挡等因素影响带来的车辆检测精度较差问题,提出一种改进区域全卷积网络的车辆检测算法(DeR-FCN)。通过特征级联的方式,跨层连接融合车辆底层细节特征和高层语义特征;使用维度分解区域提议网络,获得更加精准的区域候选框;在预测阶段采用软化非极大值抑制的方法,输出更加准确的检测结果。为验证算法的有效性,在KITTI和PASCAL VOC数据集,使用DeR-FCN算法和常用的车辆检测算法进行对比实验,实验结果表明,DeR-FCN算法的检测精度高于其它方法。There are some affecting factors such as weather,illumination,target scale and occlusion among vehicles in complex urban environments,they affect detection accuracy.To resolve the problem,a vehicle detection algorithm basing on regional full convolutional network was proposed,named as DeR-FCN.The underlying layer features and high-level semantic features of the hybrid vehicle were connected by means of feature cascading.The dimensional decomposition region proposal network was used to get a more accurate area candidate box.In the prediction stage,softer non-maximum suppression was used to output more accurate detection results.KITTI and PASCAL VOC data sets were used to verify the validity of the proposed algorithm.Experimental results show that DeR-FCN algorithm has higher detection accuracy than other methods.

关 键 词：深度学习 车辆检测 区域全卷积网络 维度分解区域提议网络 软化非极大值抑制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]