基于Transformer结构的多目标追踪算法研究综述

Overview of multi-target tracking algorithm based on Transformer structure

Transformer是一种基于编码器-解码器、完全使用自注意力机制的深度神经网络结构,目前已经成功应用于多目标追踪,性能得到大幅提升。本文首先分析了Transformer网络整体结构,归纳Transformer结构具有的优势。然后根据查询方式将基于Transformer结构的多目标追踪方法分为:基于稀疏查询的方法和基于密集查询的方法,对相关模型分析总结。最后介绍常用数据集,对比分析模型性能,指出基于Transformer结构的多目标追踪面临的挑战与未来研究方向。

As a deep neural network structure based on encoder-decoder and with the full use of self-attention mechanism,transformer has been successfully applied to the multi-target tracking now.Firstly,the overall structure of transformer network is analyzed and its advantages are summarized in the paper.Then,by the query method,the multi-target tracking methods based on transformer structure are classified into two types:sparse query and dense query,and relevant models are analyzed and summarized.Finally,the common data sets are introduced,and by comparation and analysis of the model performance,challenges and future research directions of the multi-target tracking based on the transformer structure are proposed.