摘要
针对传统基于孪生网络的目标跟踪算法面对语义背景干扰和目标遮挡导致跟踪失败的问题,提出一种基于前景概率图和动态模板的孪生网络目标跟踪算法,由基础Siamese跟踪模块、目标表征模块和模板更新模块组成。目标表征模块根据前景和背景区域的颜色分布计算区域中每个像素点与目标的相关性生成前景概率图,并掩盖到孪生网络响应图上使模型聚焦于目标,提高了模型对目标和背景的区别能力。动态模板模块将视频序列中目标外观特征矩阵分组,采用分组匹配策略进行模板更新并调整学习速率,确保了遮挡时目标模板不受污染。在主流数据集中对多种算法进行对比实验,结果表明算法精度与成功率均优于传统目标跟踪算法,相较于基准跟踪器SiamFC,精准率和成功率分别提高了11.4%与8.9%。
Focusing on the problem of the traditional object tracking algorithm based on siamese network failed in the face of semantic background interference and target occlusion,a Siamese network tracking algorithm based on prospect probability graph and dynamic template is proposed.The target representation module calculates the correlation between each pixel in the search area and the target based on the color distribution of the foreground and background area to generate the foreground probability graph,and masks it onto on the twin network response graph to make the model focus on the target,which improves the model′s ability to differentiate between the target and the background.The dynamic template module divides the target appearance feature matrix into multiple groups in the video sequence,and adopts the grouping matching strategy to update the template while adjusting the learning rate of the target template,effectively ensuring that the target template is not contaminated when the target is masked.Experimental analysis on mainstream data sets and the actual scenarios of airport flight support shows that the improved algorithm can effectively deal with complex environments such as background interference and target occlusion.
作者
薛玲祥
刘一
丁继存
XUE Ling-xiang;LIU Yi;DING Ji-cun(Qingdao Civil Aviation Cares Co.,Ltd.,Qingdao 266000,China;Civil Aviation Management Institute of China,Beijing 100000,China)
出处
《航空计算技术》
2025年第1期71-75,81,共6页
Aeronautical Computing Technique
基金
国家自然科学基金与民航基金联合重点支持项目资助(U2033214)
国家重点研发计划项目资助(2018YFB1601200)
青岛市科技惠民示范专项项目资助(22-3-7-CSPZ-19-nsh)。
关键词
目标跟踪
孪生网络
前景概率图
动态模板
时空上下文
object tracking
Siamese network
foreground probability map
dynamic template
spatiotemporal context
