机电工程技术2025，Vol.54Issue(18)：47-52,105,7.DOI:10.3969/j.issn.1009-9492.2025.18.009
一种基于Transformer的单目标跟踪算法
A Single Target Tracking Algorithm Based on Transformer
摘要
Abstract
A Transformer-based single-target tracking algorithm is proposed to address the problem of size change of tracked targets from near to far in single-target tracking.The backbone network of algorithm adopts MobileNet v2,and model complexity is reduced by removing the last four layers of network structure to improve running speed of the model.The feature fusion module uses attention mechanism to achieve the fusion and enhancement of feature map,and the detection head module applies three styles of detection heads to predict the bounding box,which are direct regression prediction box,key point regression prediction box,and mask prediction box.The branch selection module of head is designed to evaluate prediction results for two types of prediction frames,which can be evaluated before the regression prediction frame,so that it can output more ideal prediction results when tracking the target for size change.The training data set used is part of GOT-10k data set,which can achieve better tracking results on OTB data set with success rate and accuracy rate of 0.681 and 0.925,respectively.Experimental results show that the proposed algorithm can still achieve better tracking results when the tracking target undergoes a size change.关键词
目标跟踪/预测头/分支选择/大小变化/特征融合Key words
target tracking/prediction head/branch selection/size variation/feature fusion分类
信息技术与安全科学引用本文复制引用
罗阳,刘俊平,马煜,李艳豪,李龙..一种基于Transformer的单目标跟踪算法[J].机电工程技术,2025,54(18):47-52,105,7.基金项目
陕西省重点研发计划项目(2022NY-094) （2022NY-094）
陕西省教育厅专项科研项目(21JK0732) （21JK0732）
