摘要
自助收银台监控视频中商品目标之间频繁发生的遮挡现象会导致目标外观信息缺失,同时静止状态商品的运动信息无法为目标关联提供有价值的跟踪线索,导致自助收银台场景下目标跟踪困难。本研究提出一种遮挡目标全关联跟踪模型,将自助收银台场景下商品目标逐帧运动造成的遮挡现象理解为一个渐变的推理过程。模型使用基于检测的跟踪范式在目标外观信息缺失和运动信息为零时,提出遮挡率和层次信息作为遮挡目标关联的辅助信息,并借助卡尔曼滤波算法对遮挡现象推理过程中多个目标和不同轨迹之间进行关联。实验结果表明,本研究方法能够提高自助收银台场景下的商品目标跟踪精度,且遮挡率和层次信息能够有效减少目标轨迹碎片的数量,多目标跟踪精度和身份识别F1分数分别达到80.7%和80.4%,比ByteTrack模型分别提升10.6%和9.8%。
The frequent occlusion between commodity objects in self-service cashier surveillance videos can lead to the problem of missing appearance information of the objects,and the motion information of stationary products cannot provide valuable tracking clues for object association,leading to the difficulty of object tracking in self-service cashier scenarios.In this study,a fully associated tracking model for occluded objects was proposed and the occlusion phenomena caused by the frame-by-frame movement of commodity objects in the self-service cashier scenarios were interpreted as a gradual inference process.The tracking by detection(TBD)paradigm was used to propose occlusion rate and hierarchical information as auxiliary information for occlusion object association when the appearance information of the object was missing and the motion information was zero.The Kalman filtering algorithm was used to complete the association between multiple objects and different trajectories in the inference process of occlusion phenomena.The experimental results show that the proposed method can improve the tracking accuracy of commodity objects in self-service cashier scenarios and the occlusion rate and hierarchical information can effectively reduce the number of target trajectory fragments,with mutiple object tracking accuracy and identification F1 score reaching 80.7%and 80.4%respectively,10.6%and 9.8%higher than the ByteTrack model.
作者
王金富
赵建立
房胜
李哲
WANG Jinfu;ZHAO Jianli;FANG Sheng;LI Zhe(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China)
出处
《山东科技大学学报（自然科学版）》
CAS
北大核心
2024年第5期110-119,共10页
Journal of Shandong University of Science and Technology(Natural Science)
基金
山东省自然科学基金项目(ZR2022MF325)
山东省科技型中小企业创新能力提升工程项目(2021TSGC1072)。
关键词
多目标跟踪
目标检测
全关联
遮挡
自助收银台
multiple object tracking
object detection
fully-association
occlusion
self-service cashier
