用于多目标跟踪的高斯混合轨迹预测模型  

Gaussian mixed-trajectory prediction model for multi-target tracking

作　　者：丁昊[1,2] 刘运峰 石鸿凌 江小平[1,2] 孙婧[3] DING Hao;LIU Yunfeng;SHI Hongling;JIANG Xiaoping;SUN Jing(South-central Minzu University,College of Electronic and Information Engineering,Hubei Key Laboratory of Intelligent Wireless Communications,Wuhan 430074,China;South-central Minzu University,Hubei Key Laboratory of Intelligent Wireless Communications,Wuhan 430074,China;Shenzhen Institutes of Advanced Technology,Chinese Academy of Sciences,Shenzhen 518055,Guangdong China)

机构地区：[1]中南民族大学电子信息工程学院,武汉430074 [2]中南民族大学智能无线通信湖北省重点实验室,武汉430074 [3]中国科学院深圳先进技术研究院,广东深圳518055

出　　处：《中南民族大学学报（自然科学版）》2024年第6期844-850,共7页Journal of South-Central Minzu University(Natural Science Edition)

基　　金：国家重点研发计划资助项目(2020YFC1522900);中央高校攻关计划专项资金资助项目(CZT20002);广东省基础与应用基础研究基金资助项目(2021A1515011913)。

摘　　要：针对视频中目标之间的遮挡会增加轨迹关联难度的问题,提出了一种高斯混合轨迹预测模型来改善遮挡情况下的关联效果.该模型采用Transformer结构,从对象的历史运动信息中提取更长的时间依赖性,以便更准确地预测跟踪对象在未来的轨迹分布;同时,将模型的输出设置为高斯混合分布,采用多个高斯分布相加来定义跟踪对象在未来的轨迹分布,以利用轨迹分布来计算检测目标和跟踪对象之间的空间相似性.实验结果表明:该模型能够改善跟踪性能,在MOT17数据集上HOTA达到了56.3%.To address the challenge of increased difficulty in tracking object associations due to occlusions among targets in videos,a Gaussian Mixture Trajectory Prediction model is proposed to enhance the association performance under occlusion scenarios.This model employs a Transformer architecture to capture longer temporal dependencies from the historical motion information of objects,thereby accurately predicting the future trajectory distribution of tracked objects.Additionally,the model sets the output as a Gaussian mixture distribution,which combines multiple Gaussian distributions to define the future trajectory distribution of the tracked objects.This enables the utilization of trajectory distributions to calculate the spatial similarity between detection targets and tracked objects.Experimental results demonstrate that this model improves tracking performance,achieving a HOTA score of 56.3%on the MOT17 dataset.

关 键 词：多目标跟踪 Transformer结构 运动预测 高斯混合模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]