摘要
针对前方运动车辆复杂场景下的跟踪精度较低的问题,文中将庞大的VGG-M网络模型应用到实时跟踪中,并结合在线观测模型,实现对前方车辆稳定精准的跟踪。通过改进样本生成方案,优化网络训练集,提高了网络训练效率。采用自适应更新模型,可根据目标轮廓的高宽比、内部信息熵和跟踪的尺度置信度实时调节网络更新频率。实验结果表明,在线VGG-M跟踪模型比传统的车辆跟踪方法的性能有明显的改善。
Aiming at the low accuracy of front moving vehicle tracking in complex scenes,the huge VGG-M network model is applied to real-time tracking,and the online observation model is used to achieve stable and accurate tracking of front vehicles.By improving the sample generation scheme and optimizing the network training set,the efficiency of network training is enhanced.Furthermore,with adaptive update model adopted,the network update frequency can be adjusted in real time according to the aspect ratio of target profile,internal information entropy and the confidence of tracking scale.Experimental results show that the online VGG-M tracking model achieves better performance than the traditional vehicle tracking methods.
作者
刘国辉
张伟伟
吴训成
宋晓琳
许莎
温培刚
Liu Guohui;Zhang Weiwei;Wu Xuncheng;Song Xiaolin;Xu Sha;Wen Peigang(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201600;;Hunan University,Laboratory of Advanced Design and Manufacturing for Vehicle Body,Changsha 410082)
出处
《汽车工程》
EI
CSCD
北大核心
2019年第1期57-63,共7页
Automotive Engineering
基金
国家自然科学基金(51675324
51575169
51805312)
上海高校青年教师培养计划(ZZGCD15102)
上海工程技术大学校启科研项目(2016-19)
上海工程技术大学研究生科研创新项目(16KY0602)
第八批(2017年)"上海高校教师产学研计划"(A3-0100-17-SDJH337)
上海市青年科技英才扬帆计划(18YF1409400)资助
关键词
深度学习
前车跟踪
在线观测模型
网络自适应更新模型
deep learning
front vehicle tracking
online observation model
network adaptive update model
