面向行为理解的多人姿态估计研究综述  

A Review of Research on Multi-person Pose Estimation for Behavior Understanding

作　　者：王同喜 胡彩平 张航川 陈彭 WANG Tongxi;HU Caiping;ZHANG Hangchuan;CHEN Peng(Jinling Institute of Technology,Nanjing 211169,China)

机构地区：[1]金陵科技学院计算机工程学院,江苏南京211169

出　　处：《金陵科技学院学报》2025年第1期33-42,共10页Journal of Jinling Institute of Technology

基　　金：江苏省重点研发计划项目(BE2022077);金陵科技学院高层次人才科研启动项目(jit-rcyj-202102)。

摘　　要：针对多人姿态估计技术的发展现状、主流方法及其在行为理解中的关键技术融合展开了综述。首先,对从单人到多人姿态估计的技术演进路径进行了梳理,并对自顶向下、自底向上和Transformer-based这三类框架的优缺点展开了对比分析。其次,探讨了时序建模、多模态融合以及层级化行为建模等关键技术对行为理解能力的提升作用。最后,通过对主流数据集和实验结果的归纳总结,揭示了当前方法在复杂场景下的性能瓶颈,并提出未来需进一步攻克的难题,如遮挡鲁棒性、实时性优化和多尺度融合等,为面向行为理解的多人姿态估计研究提供了技术参考与发展方向。A review was conducted on the current development status,mainstream methods,and key technology integration in behavior understanding of multi-person pose estimation technology.Firstly,the technical evolution path from single-person to multi-person pose estimation was sorted out,and a comparative analysis was conducted on the advantages and disadvantages of three types of frameworks:top-down,bottom-up,and Transformer-based.Secondly,the role of key technologies such as time series modeling,multimodal fusion,and hierarchical behavior modeling in improving behavior understanding capabilities was explored.Finally,through the summary of mainstream datasets and experimental results,the performance bottlenecks of current methods in complex scenarios were revealed,and future challenges that need to be further overcome were proposed,such as occlusion robustness,real-time optimization,and multi-scale fusion.This paper provides technical references and development directions for research on multi-person pose estimation for behavior understanding.

关 键 词：深度学习 图像识别 多人姿态估计 行为理解 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]