基于工地场景的深度学习目标跟踪算法  

Deep Learning Target Tracking Algorithm Based on Construction Site Scene

作　　者：马少雄 邱实[3] 唐颖[4] 张晓[5] MA Shao-xiong;QIU Shi;TANG Ying;ZHANG Xiao(Xi’an University of Technology,Xi’an,Shaanxi 710048,China;Shaanxi Railway Institute,Weinan,Shaanxi 714000,China;Key Laboratory of Spectral Imaging Technology CAS,Xi’an Institute of Optics and Precision Mechanics,Chinese Academy of Sciences,Xi’an,Shaanxi 710119,China;Chengdu University of Technology,Chengdu,Sichuan 610059,China;Northwest University,Xi’an,Shaanxi 710127,China)

机构地区：[1]西安理工大学,陕西西安710048 [2]陕西铁路工程职业技术学院铁成(创新)学院,陕西渭南714000 [3]中国科学院西安光学精密机械研究所中科院光谱成像技术重点实验室,陕西西安710119 [4]成都理工大学信息与科学学院电子工程系,四川成都610059 [5]西北大学信息科学与技术学院,陕西西安710127

出　　处：《电子学报》2020年第9期1665-1671,共7页Acta Electronica Sinica

基　　金：中国科学院“西部之光”项目(No.XAB2016B23)。

摘　　要：针对施工现场环境复杂,难以高效管理的问题.提出了基于工地场景的深度学习目标跟踪算法,辅助施工顺利进行.根据工地现场目标的连续性,构建增强群跟踪器,提升目标成功跟踪的概率.然后从滑动窗口、Stacked Denoising Auto Encoder(SDAE)和Support Vector Machine(SVM)三方面组建深度检测器.在滑动窗口方面:从梯度角度建立模型实现窗口自适应.在SDAE算法方面:构建反向算法微调网络参数.优化SVM算法降低跟踪时目标漂移和跟踪失败的概率,最终实现目标高精度跟踪.通过实验表明本文提出的算法可有效对目标进行跟踪,实现动态管理.Construction site is difficult to be effectively managed owing to its complex environment.A deep learning target tracking algorithm based on construction site scene is proposed to assist the construction progress.Firstly,according to the continuity of the target in the site scene,the enhanced group tracker is constructed to improve the successful probability of target tracking.Then,the depth detector is constructed with sliding window,stacked denoising auto encoder(SDAE)and support vector machine(SVM).Sliding window:a model is built from the gradient angle to realize window adaption.SDAE algorithm:the reverse algorithm is built to fine-tune network parameters.Optimized SVM algorithm reduces the probability of target drift and tracking failure.Finally,high precision tracking is achieved.Experiments show that the proposed algorithm can track the target effectively and realize dynamic management.

关 键 词：工地场景 深度学习 目标跟踪 增强群滤波器 SDAE SVM 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]