摘要
针对单一图像源下目标跟踪鲁棒性差和跟踪精度低的问题,论文提出一种基于卷积神经网络(CNN)特征的RGB-T目标鲁棒性跟踪算法。首先,采用分层CNN特征对RGB图像和热红外图像进行编码。其次,基于SiamDW跟踪框架对目标进行跟踪。然后根据短时间内的跟踪结果对每个CNN特征的结果进行自适应融合并定位。最后,将RGB图像和热红外图像的结果进行融合并定位。实验表明,与现有的孪生跟踪算法相比,该算法在中心位置偏差和重叠率上表现更优,且在复杂情况下鲁棒性更好。
Aiming at the problem of poor robustness and low tracking accuracy of target tracking under a single image source,this paper proposes a robust tracking algorithm of RGB-T target based on Convolutional Neural Network(CNN)features.Firstly,layered CNN features are used to encode RGB images and thermal infrared images.Secondly,targets are tracked based on SiamDW tracking framework.Then according to the results of tracking in a short period of time the results of each CNN features for adaptive fusion and locate in the end,the RGB image and the result of thermal infrared image fusion and locate experiments show that compared with the existing siamese tracking algorithm,this algorithm has better performance in a central location deviation and overlap rate and better robustness in complex situations.
作者
刘莲
李福生
LIU Lian;LI Fushen(School of Automation Engineering,University of Electronic Science and Technology of China,Chengdu 611730;Yangtze Delta Region Institute,University of Electronic Science and Technology of China,Huzhou 313000)
出处
《计算机与数字工程》
2024年第2期432-435,共4页
Computer & Digital Engineering
基金
国家自然科学基金项目“基于基本参数法和智能反演算法的XRF光谱快速建模理论研究”(编号:62075028)资助。