结合通道选择与位置优化的目标跟踪算法  

Object Tracking Algorithm Combining Channel Selectionand Location Optimization

作　　者：吴捷[1,2] 马小虎[2] WU Jie;MA Xiaohu(School of Information Technology,Taizhou Polytechnic College,Taizhou 225300,China;Computer Science and Technology College,Suzhou University,Suzhou 215006,China)

机构地区：[1]泰州职业技术学院信息技术学院,江苏泰州225300 [2]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《火力与指挥控制》2022年第8期25-31,共7页Fire Control & Command Control

基　　金：国家自然科学基金(61402310);江苏省自然科学基金(BK20141195);泰州职业技术学院重点科研基金资助项目(1821819039)。

摘　　要：为了解决TADT等算法在目标发生形变及低分辨率等复杂场景下易发生跟踪漂移的问题,在Siamese孪生网络框架下,结合通道选择与位置优化提出一种新型目标跟踪算法。利用VGG16网络中的Conv4-3层进行目标初次定位,进而利用峰值旁瓣比对跟踪结果进行可信度判断,将结果分为高可信、待优化和低可信3种类别。对于待优化类别,利用Conv3-3层深度特征在目标初次定位基础上进行精确再定位。对于低可信类别,提取Conv5-1层深度特征进行目标重定位。在OTB-50及OTB-100数据集上进行了综合评估并与当前的部分先进算法进行了比较,实验结果表明,算法具有最佳跟踪精度且相对TADT算法有了全面提升,跟踪速度达到24.8 FPS。In order to solve the tracking drift problem of TADT and other algorithms in complex scenes such as target deformation and low resolution,a new object tracking algorithm based on channel selection and location optimization under the framework of Siamese network is proposed.First,the object is initially located by using the conv4-3 layer of VGG16 network,and then the reliability of the tracking results is judged by using the peak side lobe ratio.The results are divided into three categories:high reliability,to be optimized and low reliability.For the target to be optimized,the deep feature of conv3-3 layer is used to locate the target accurately based on the initial location of target.For the case of low reliability,the deep feature of conv5-1 layer is extracted to relocate the target.The comprehensive evaluation is carried out on the base of OTB-50 and OTB-100 datasets,and the comparison is made with some of the current advanced algorithms.The experimental results show that the proposed algorithm has the best tracking accuracy and a comprehensive improvement compared with TADT algorithm,and the tracking speed reaches 24.8 FPS.

关 键 词：通道选择 位置优化 目标跟踪 深度特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]