SiamMT:基于自适应特征融合机制的可修正RGBT目标跟踪算法  

SiamMT:A modifiable RGBT target tracking algorithm based on adaptive feature fusion mechanism

作　　者：齐咏生[1,2,3] 姜政廷 刘利强 苏建强[1,2] 张丽杰 QI Yong-sheng;JIANG Zheng-ting;LIU Li-qiang;SU Jian-qiang;ZHANG Li-jie(College of Electric Power,Inner Mongolia University of Technology,Hohhot 010080,China;Intelligent Energy Technology and Equipment Engineering Research Centre of Colleges and Universities in the Inner Mongolia Autonomous Region,Hohhot 010080,China;Engineering Research Center of the Ministry of Education for Large Scale Energy Storage Technology,Hohhot 010080,China)

机构地区：[1]内蒙古工业大学电力学院,呼和浩特010080 [2]内蒙古自治区高等学校智慧能源技术与装备工程研究中心,呼和浩特010080 [3]大规模储能技术教育部工程研究中心,呼和浩特010080

出　　处：《控制与决策》2025年第4期1312-1320,共9页Control and Decision

基　　金：国家自然科学基金项目(62363029);内蒙古科技计划项目(2021GG0164);内蒙古自然科学基金项目(2022MS06018,2021MS06018);高校院所协同创新项目(XTCX2023-16)。

摘　　要：针对传统RGBT目标跟踪算法网络精确度低、鲁棒性差,以及在目标尺度变化大和长时跟踪过程中存在目标丢失无法找回等问题,提出一种新的基于自适应特征融合机制的可修正RGBT目标跟踪算法.首先,引入一种特征层与模态间双自适应融合机制,充分利用两模态间的互补信息,增强RGB与红外特征的跨模态融合;然后,设计一种后端时序约束回归模块,利用上一帧信息对IOU计算以及边界框回归进行约束,有效减少相似物干扰;最后,提出一种基于元学习的在线模板更新机制,对回归阶段得分较高的模板图像进行更新存储,解决长时跟踪中累计误差和目标难以找回问题.采用权威的目标跟踪数据集GTOT、RGBT234和VOT-RGBT2019进行算法验证,所提出方法均可取得极具竞争力的结果.将算法移植到嵌入式设备Jetson Xavier NX上进行性能测试,实验结果表明:所提出算法运行速度可达到29帧/s,相比于当前流行的多种RGBT算法,具有更为全面的跟踪性能,且能够有效解决相似物干扰、目标丢失难找回等问题.In order to solve the problems of low network accuracy and poor robustness of traditional RGBT target tracking algorithms,as well as the problems of target loss and unretrieval in the process of large target scale change and long-term tracking,a new modifiable RGBT target tracking algorithm based on adaptive feature fusion mechanism is proposed.Firstly,an adaptive dual-modal fusion module is introduced to make full use of the complementary information between the two modalities to enhance the cross-modal fusion of RGB and infrared features.Then,a backend timing constraints regression module is designed,using the information of the previous frame to constrain the IOU calculation and bounding box regression,which effectively reduces the interference of similarities.Finally,an online template update mechanism based on meta-learning is proposed to update and store the template images with high scores in the regression stage,so as to solve the problems of cumulative error and difficult target recovery in long-term tracking.Using the authoritative object tracking datasets GTOT,RGBT234 and VOT-RGBT2019 for algorithm verification,the proposed method can achieve very competitive results.The algorithm is transplanted to the embedded device Jetson Xavier NX for performance testing.The testing results show that the proposed algorithm runs at a speed of 29 frames/s,which has more comprehensive tracking performance than the current popular RGBT algorithms,and can effectively solve the problem of similar interference,problems such as the difficulty of retrieving the lost target.

关 键 词：目标跟踪 孪生网络 RGBT 元学习 特征融合 

分 类 号：TP394.41[自动化与计算机技术—计算机应用技术]