摘要
为了精确捕捉人体在各种复杂动作中的姿态变化,并减少动作轮廓捕捉的误判和漏判情况,研究一种基于条件GAN的人体复杂动作轮廓智能捕捉方法。构建人体复杂动作图像前景模型,利用该前景模型去除人体复杂动作图像内的干扰背景,获得其前景图像。将人体复杂动作前景图像输入到条件GAN模型内,该模型使用生成器,依据叠加条件提取人体复杂动作前景图像特征,并生成人体复杂动作轮廓图像,将该轮廓图像输入到判别器内,判别器对生成器生成的人体复杂动作轮廓图像进行判别,输出其判别值;同时,生成器依据该判别值对人体复杂动作轮廓图像进行调整和优化。生成器和判别器不断对抗,当条件GAN模型的损失函数达到最小时,生成器输出最终人体复杂动作轮廓图像,实现人体复杂动作轮廓智能捕捉。实验结果表明,所提方法可有效捕捉人体复杂动作轮廓,且智能捕捉的人体复杂动作轮廓周长与其实际轮廓周长差异较小。
In order to accurately capture the posture changes of the human body in various complex movements and reduce the misjudgment and omission of motion contour capture,a conditional GAN based intelligent capture of complex human motion contour is studied.A foreground model for complex human motion images is constructed,and the foreground model is used to remove interfering backgrounds within the complex human motion images,to obtain their foreground images.The complex human motion foreground image is input into the conditional GAN model.This model can use the generator to extract the features of the complex human motion foreground image according to the superposition conditions,and generate the complex human motion contour image.The contour image is input into the discriminator,which can judge the complex human motion contour image generated by the generator and output its discrimination value.The generator can adjust and optimize the complex human motion contour image according to the discrimination value.The generator and discriminator constantly compete,and when the loss function of the conditional GAN model reaches its minimum,the generator can output the final complex human motion contour image,achieving the intelligent capture of complex human motion contours.The experimental results show that the proposed method can effectively capture complex human motion contours,and the difference between the circumference of the intelligently captured complex human motion contour and its actual contour circumference is small.
作者
王鹏博
刘菡
WANG Pengbo;LIU Han(Yunnan Minzu University,Kunming 650500,China)
出处
《现代电子技术》
北大核心
2024年第24期172-176,共5页
Modern Electronics Technique
关键词
条件GAN
人体复杂动作
轮廓图像
智能捕捉
前景模型
生成器
判别器
conditional GAN
complex human motion
contour image
intelligent capture
prospect model
generator
discriminator
作者简介
王鹏博(1999-),男,河南沈丘人,在读硕士研究生,研究方向为计算机算法、动作识别;刘菡(1989-),女,河南周口人,在读博士研究生,助理研究员,研究方向为运动图像处理。
