改进转换模型预测网络的外观变化下目标跟踪方法  

Improved ToMP network appearance changes under target tracking method

作　　者：陈仁祥 何家乐 杨黎霞 余腾伟 梁栋 Chen Renxiang;He Jiale;Yang Lixia;Yu Tengwei;Liang Dong(Chongqing Engineering Laboratory for Transportation Engineering Application Robot,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing University of Science&Technology,Economics and Finance College,Chongqing 401331,China)

机构地区：[1]重庆交通大学交通工程应用机器人重庆市工程实验室,重庆400074 [2]重庆科技学院经济与金融学院,重庆401331

出　　处：《仪器仪表学报》2024年第9期92-100,共9页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金项目(52475548);重庆市教委科学技术研究项目(KJZD-M202200701);重庆市自然科学基金创新发展联合基金项目(CSTB2023NSCQ-LZX0127);重庆市研究生联合培养基地项目(JDLHPYJD2021007);重庆市专业学位研究生教学案例库项目(JDALK2022007);重庆市自然科学基金项目(CSTB2023NSCQ-MSX0177)资助。

摘　　要：针对跟踪过程中因目标发生外观变化导致跟踪效果不佳的问题,提出改进转换模型预测(ToMP)网络的外观变化下目标跟踪方法。首先,在ToMP网络基础上添加目标外观状态判别模块,利用级联长短时记忆网络(LSTM)输出目标外观状态判别信息;其次,优化网络在线样本储存判据,在置信度分值判别基础上添加正常外观判别信息,利用可靠样本优化模型权重,增强网络对目标的分类能力;然后,改进网络外观变化时在线样本利用机制,结合最新样本优化模型权重,提高网络对新外观目标的分类效果;最后,利用中心点轨迹预测位置对网络生成目标响应分数进行加权,增强目标特征映射效果的同时减少相似物体干扰以稳定跟踪目标。在公共数据集上准确率分别达到93.9%和68.9%,优于其他方法,并利用机器人进一步证实了可行性。To address the issue of poor tracking performance caused by target appearance changes during tracking,a target tracking method under appearance change based on improved ToMP network is proposed.Firstly,a target appearance state discrimination module was added to the ToMP network,using a cascaded Long Short-Term Memory(LSTM)network to output the target appearance state discrimination information.Secondly,the online sample storage criterion of the network was improved by adding the normal appearance discrimination information to the confidence score-based evaluation.This optimizes the model weights using reliable samples,enhancing the network′s classification ability for the target.Then,the mechanism for utilizing online samples during appearance changes was improved,updating the model weights with the latest samples to enhance classification performance for newly appeared targets.Finally,a center point trajectory prediction was used to weight the target response score generated by the network,improving the target feature mapping while reducing interference from similar objects,thereby stabilizing target tracking.The accuracy on public datasets reached 93.9%and 68.9%,respectively,outperforming other methods and feasibility is further confirmed through robotics experiments.Aiming at the problem that the tracking effect is not good due to the appearance change of the target in the tracking process,a target tracking method under appearance change based on improved ToMP network is proposed.Firstly,a target appearance state discrimination module was added to the ToMP network,and the cascaded Long Short-Term Memory(LSTM)network was used to output the target appearance state discrimination information.Secondly,the online sample storage criterion of the network was improved,the normal appearance discrimination information was added on the basis of the confidence score discrimination,and the reliable samples were used to optimize the model weight to enhance the classification ability of the network for the target.Then,the

关 键 词：目标跟踪 改进转换模型 连续状态判别 预测距离加权 

分 类 号：TH39[机械工程—机械制造及自动化]