首页|基于机器视觉的运动动作多特征识别算法研究

基于机器视觉的运动动作多特征识别算法研究

扫码查看
针对传统特征设计方法的兴趣点泛化能力弱、可迁移性差等问题,研究首先提出一种基于多特征学习的卷积神经网络,然后引入跨模态训练对光流图改进,最后使用运动激励和时间聚合模块进行优化,最终得到基于机器视觉的舞蹈运动动作多特征识别算法.研究结果显示,在FolkDance舞蹈数据集与AIST++舞蹈数据集中研究并提出的算法融合准确率分别为98.1%与74.3%,且在仅需1.73 s就可对舞蹈动作实现精准识别.综上所述,研究提出的方法能极大减少人工与时间成本,对复杂的舞蹈运动视频中的舞蹈动作能实现精准识别,在实际应用中具有更强的适用性.
Research on Multiple Feature Recognition Algorithm for Motion Actions Based on Machine Vision
To solve the problems of weak generalization ability and poor mobility of interest points in traditional feature design methods,the research first proposes a Convolutional neural network based on multi feature learning,then introduces cross modal train-ing to improve the optical flow graph,and finally uses the motion excitation and time aggregation module to optimize,and finally ob-tains the multi feature recognition algorithm of dance movement based on machine vision.The research results show that the fusion ac-curacy of the algorithms proposed in the FolkDance and AIST++dance datasets is 98.1%and 74.3%,respectively,and precise rec-ognition of dance movements can be achieved in just 1.73 seconds.In summary,the proposed method can greatly reduce the cost of labor and time,achieve accurate recognition of dance movements in complex dance motion videos,and have stronger applicability in practical applications.

machine visionmulti feature learningdance sportshuman motion recognitioncross modal pre trainingincentive module

张楠

展开 >

陕西铁路工程职业技术学院,陕西渭南 714000

机器视觉 多特征学习 舞蹈运动 人体动作识别 跨模态预训练 激励模块

陕西省高等学校学生工作研究课题(2022)

2022XKT65

2024

自动化与仪器仪表
重庆工业自动化仪表研究所,重庆市自动化与仪器仪表学会

自动化与仪器仪表

CSTPCD
影响因子:0.327
ISSN:1001-9227
年,卷(期):2024.(3)
  • 12