摘要
多目标跟踪中的JDE算法首次将目标检测与重识别进行联合学习,极大提升了跟踪速度,但由于复杂背景干扰和遮挡导致跟踪准确度下降。为了解决跟踪速度与准确度的平衡问题,本文提出了SAM-JDE,该模型融合了SimAM注意力机制、多尺度融合等思想,通过增强特征提取能力提高目标跟踪的准确性。使用CIoU_Loss作为回归损失函数,通过准确地构建目标框和预测框之间的位置关系来提升定位精度。关联匹配部分使用卡尔曼滤波预测运动信息,匈牙利匹配算法完成时序维度上的目标关联。在MOT16-test数据集上进行测试,MOTA达到66.4%,跟踪速度为20.6 FPS,在保证实时性的基础上跟踪准确度较JDE算法提升2.3%,较好地优化了准确度与速度的平衡问题。
JDE algorithm in multi-object tracking jointly learns target detection and re-identification for the first time,which greatly improves the tracking speed.However,the tracking accuracy is reduced due to the poor tracking effect caused by complex background interference and occlusion processing.In order to balance the tracking speed and accuracy,SAM-JDE is proposed in this paper.This model integrates SimAM attention mechanism,multi-scale fusion and other ideas to improve the accuracy of target tracking by enhancing the ability of feature extraction.CIoU_Loss is used as the regression loss function to improve the positioning accuracy by accurately building the position relationship between the target box and the prediction box.In the association matching part,Kalman filtering is used to predict the motion information,and the Hungarian matching algorithm completes the target association in the time series dimension.Testing on MOT16-test dataset shows that MOTA reaches 66.4%and tracking speed is 20.6 FPS.On the basis of ensuring real-time performance,tracking accuracy is 2.3%higher than JDE algorithm,which better optimizes the balance between accuracy and speed.
作者
程之星
杨帆
Cheng Zhixing;Yang Fan(School of Electronics Information Engineering,Hebei University of Technology,Tianjin 300401,China)
出处
《电子测量技术》
北大核心
2023年第17期94-101,共8页
Electronic Measurement Technology
基金
国家重点研发计划智能机器人专项(2019YFB1312102)
河北省自然科学基金(F2019202364)项目资助。
关键词
机器视觉
多目标跟踪
注意力机制
实时跟踪
machine vision
multi-object tracking
attention mechanism
real time tracking
