结合超像素的动态单目标追踪方法  

Single Video Object Tracking Method Combined with Superpixels

作　　者：程金伟 胡大裟[1] 蒋玉明[1] CHENG Jinwei;HU Dasha;JIANG Yuming(College of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《现代计算机》2021年第15期18-22,共5页Modern Computer

摘　　要：随着卷积神经网络在目标追踪领域的应用愈加广泛,目标追踪算法也面临越来越高的准确率要求。基于孪生网络结构的目标追踪算法使用卷积神经网络进行目标特征提取,将追踪目标特征与后续帧中的候选目标特征进行匹配,以预测目标位置并生成对应掩膜,通过超像素分割算法对候选目标位置进行分割,进一步优化卷积神经网络输出的目标掩膜,提高目标追踪算法的准确率。With the increasing application of convolutional neural networks in the field of object tracking,single object tracking algorithms are also facing higher requirements in accuracy.In order to perform template matching between object in the first frame and candidates in latter frames,siamese network based tracking algorithms use convolutional neural network to extract features to predict the potential object posi⁃tion and generate corresponding mask.Further,the using of superpixel method in corresponding candidate positions optimizes masks out⁃put by convolutional neural network,leading to higher accuracy of object tracking.

关 键 词：目标追踪 孪生网络 深度学习 机器视觉 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]