基于元素相乘结构的实时多目标跟踪算法  

Real-time multi-object tracking algorithm based on star operation

作　　者：周畅达 杨帆 Zhou Changda;Yang Fan(School of Electronics Information Engineering,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学电子信息工程学院,天津300401

出　　处：《电子测量技术》2025年第3期43-51,共9页Electronic Measurement Technology

基　　金：石家庄市科技合作专项重大项目(SJZZXA23005)资助。

摘　　要：多目标跟踪算法中的FairMOT提出了平衡检测和重识别分支的均衡学习策略,有效的平衡了目标检测和重识别两大任务,是目前单阶跟踪段范式算法中最优的算法,但由于DLA34骨干网络的特征提取能力有限,面对实际应用场景中复杂的跟踪场景时,往往会因为出现漏检和误跟等现象导致模型的跟踪效果下降。为了有效的提升模型骨干网络的特征提取能力,本文针对此问题设计了基于元素相乘结构的深度聚合骨干网络,提出了FairMOT-Star算法。该算法利用了元素相乘结构带来的隐藏维度提升原理,实现了简洁高效的目标特征提取。同时使用EIoU_Loss作为检测框回归任务的回归损失函数,更加精准的描述了检测框和真实框之间的位置和形状关系,提升了检测框的预测精度。匹配关联部分使用卡尔曼滤波算法预测目标的运动信息,匈牙利算法完成时序维度上前后帧目标和轨迹的关联匹配。在MOT16数据集上进行了实验测试,MOTA精度达到了86.0%,模型的权重参数量为19.59 M,相比于FairMOT模型参数量减少9.7%的同时,MOTA精度提升了3.5%,较好的优化了FairMOT算法的计算参数量和跟踪精度。FairMOT,a multi-object tracking algorithm,proposes a balanced learning strategy between the detection branch and the re-identification branch,effectively balancing the tasks of object detection and re-identification,thereby improving tracking accuracy.However,due to the limited feature extraction capability of its DLA34 backbone network,the model′s tracking performance often declines in complex real-world scenarios,leading to missed detections and incorrect tracking.To enhance the backbone network′s feature extraction capability,this paper designs a deep aggregated backbone network based on an element-wise multiplication structure and proposes the FairMOT-Star algorithm.This algorithm leverages the principle of hidden dimension enhancement brought by the element-wise multiplication structure to achieve concise and efficient object feature extraction.Additionally,EIoU_Loss is used as the regression loss function for the bounding box regression task,more precisely describing the positional and shape relationships between detection boxes and ground truth boxes,thus improving prediction accuracy.In the matching and association part,the Kalman filter algorithm predicts target motion information,and the Hungarian algorithm associates and matches targets and trajectories across frames in the temporal dimension.Experimental tests on the MOT16 dataset achieved an MOTA accuracy of 86.0%.The model′s weight parameters amount to 19.59 M,reducing parameter count by 9.7%compared to the FairMOT model,while increasing MOTA accuracy by 3.5%,effectively optimizing the computational parameters and tracking accuracy of the FairMOT algorithm.

关 键 词：机器视觉 多目标跟踪 特征提取 FairMOT 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN98[自动化与计算机技术—计算机科学与技术]