基于时空Transformer的视觉目标跟踪算法  

Visual object tracking algorithm based on spatio-temporal transformer

作　　者：武晓军 陈怡丹 冯丽萍[1] 宋长伟 何德清 WU Xiaojun;CHEN Yidan;FENG Liping;SONG Changwei;HE Deqing(Xinzhou Normal University,Xinzhou 034000,China;Henan Open University,Zhengzhou 450046,China;College of Information and Management Science,Henan Agricultural University,Zhengzhou 450003,China;School of Computer Science&Technology,Huazhong University of Science and Technology,Wuhan 430074,China)

机构地区：[1]忻州师范学院,山西忻州034000 [2]河南开放大学,河南郑州450046 [3]河南农业大学信息与管理科学学院,河南郑州450003 [4]华中科技大学计算机科学与技术学院,湖北武汉430074

出　　处：《传感器与微系统》2025年第3期152-155,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61902139);山西省基础研究计划项目(202203021211116);忻州师范学院项目(2021KY16)。

摘　　要：视觉目标跟踪中,由于目标移动速度不同,连续帧对时空邻域的贡献程度也不同。为学习视频帧对邻域信息的贡献,结合自注意力机制学习不同帧的权重大小,提出了一种基于时空Transformer的视觉目标跟踪方法。该算法主要通过关联多帧特征,并在时域上进行信息聚合。首先,将图像通过空间Transformer编码器(STE)对空间特征进行编码。然后,通过时空Transformer解码器(STD)模块在时间维度上聚合帧间信息,以捕获时间和空间的全局上下文信息。最后,在LaSOT、GOT—10k等主流数据集进行测评。实验结果表明:算法在精度、成功率及其他评价指标上取得了一定程度的提升。In visual object tracking,different continuous contribute degrees to the spatio-temporal neighborhood are different due are different object moving speeds.In order to learn the contribution of video frames to the neighborhood information,a visual object tracking method based on spatio-temporal Transformer is proposed by combining the self-attention mechanism to learn the weight magnitude of different frames.The algorithm mainly works by associating multi-frame features and aggregating the information in the time domain.The images are firstly encoded by spatial transformer enconder(STE)spatial features,and then inter-frame information is aggregated in the temporal dimension by spatial-temporal transformer deconder(STD)module to capture the global contextual information in time and space.Finally,it is measured in mainstream datasets,and the experimental results show that the algorithm achieves a certain degree of improvement in precision,success rate,and other evaluation metrics.

关 键 词：视觉跟踪 TRANSFORMER 时空特征 自注意力 特征编码 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]