一种视频中时空动作检测的交互关系建模方法  

An Interactive Relationship Modeling Method for Spatio-temporal Action Detection in Video

作　　者：李飞[1] LI Fei(School of Computer and Artificial Intelligence,Southwest Jiaotong University,Chengdu Sichuan 611756,China)

机构地区：[1]西南交通大学计算机与人工智能学院,四川成都611756

出　　处：《信息与电脑》2022年第8期6-8,共3页Information & Computer

摘　　要：目前,在时空动作检测任务中,很少有方法明确提出使用优化动作交互关系特征来提升检测的准确性。笔者提出基于多头注意力的交互特征增强方法,包括人物视觉、人的时序、全局背景和空间位置的特征增强,采用串行融合策略连接增强块,有效提升了时空动作检测的准确性。模型在自建Unload-Actions数据集上的帧级mAP达到27.93%,并在消融实验中证明了交互特征增强块和串行融合策略的有效性。The current spatiotemporal action detection methods have problems in the feature representation and fusion strategy of interaction relationships in actions. This paper proposes an interactive feature enhancement method based on multi-head attention, including feature enhancement of character vision, human timing, global background, and spatial location. A serial fusion strategy is used to connect enhancement blocks, which greatly improves the accuracy of spatiotemporal action detection. The model achieves 27.93% frame-level mAP on the self-built Unload-Actions dataset and demonstrates the effectiveness of the interactive feature enhancement block and serial fusion strategy in ablation experiments.

关 键 词：时空动作检测 交互关系建模 多头注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]