基于机器视觉的婴儿全身运动质量智能评估

Three-dimensional Material Intelligent Assessment of Infant Whole-body Movement Quality Based on Machine Vision

针对婴儿全身运动质量评估问题,基于姿势识别对婴儿运动特征进行提取与分析,提出基于ResNet和反卷积层的婴儿姿势热力图识别模型,平均识别率达到86.9%;利用婴儿的二维姿势坐标,建立基于DenseNet的3D人体姿势识别模型,使用1D卷积网络及1D连接层,实现婴儿的2D姿势坐标到3D姿势坐标的升维推算;使用四元数作为空间向量表示方式,对婴儿主要肢体运动的角度、角速度、角加速度进行提取,并提出基于支持向量机(support vector machine,SVM)的由婴儿肢体运动角度特征进行判定的婴儿异常行为识别模型。针对模型参数过多的问题,在保证模型整体识别率的情况下,使用主成分分析的方式对模型进行特征降维,提高整体识别速度,结果表明,将维度由18维度降低至8维度后,整体运行时间减少近50%,且对于不同的视频样品均能正确分类。

Aiming at the problem of infant whole-body motion quality assessment,by extracting and analyzing infant motion features based on posture recognition,an infant heat map recognition model based on ResNet and inverse convolution layer was proposed,with an average recognition rate of 86.9%.Using infant s 2D pose coordinates,a 3D human pose recognition model based on DenseNet was established,using 1D convolutional network and 1D connection layer.The 2D pose coordinates of the infant were used to up-dimension the 3D pose coordinates.The angles,angular velocities,and angular accelerations of the infant s main limb movements were extracted using quaternions as the spatial vector representation,and an support vector machine(SVM)-based infant abnormal behavior recognition model determined by the angular characteristics of the infant s limb movements was proposed.To address the problem of too many model parameters,the overall recognition rate of the model was guaranteed,and the overall recognition speed was improved by using principal component analysis for feature dimensionality reduction.The results show that after reducing the dimensionality from 18 to 8,the overall running time is reduced by nearly 50%,and different video samples can be correctly classified.