基于机器学习的人体动作局部特征点识别仿真

Simulation of Human Body Local Feature Points Recognition Based on Machine Learning

为了解决对人体动作局部特征点的识别误差率较高的问题,提出一种基于机器学习的人体动作局部特征点识别方法。首先利用人体在时空状态下的差别及运动频率变化,构建多尺度的局部时空领域特征。以目标之前状态为基础,通过卡尔曼滤波计算法对人体关节位置评估,并对之后的状态做误差最优估计,以此构建人体行为数学模型。利用小波转换函数构建神经网络模型,将之前所提取出来的人体动作特征点参数作为输入神经元,输入进神经网络内进行训练。然后利用训练完毕的神经网络对测试集分类进行定量识别,并确定识别精准度指标。仿真结果表明,所提方法的相对于传统方法提取人体动作数据多,分类全面,证明了所提方法识别准确度较高、误差率较低。

In order to solve the problem of high error rate in the recognition of local feature points of human action, this article presented a method for recognizing local feature points of human motion based on machine learning. Based on the difference in the human body under the spatio-temporal state and the change in motion frequency, the multi-scale local space-time domain characteristics were constructed. Based on the previous state of the target, the position of the human joint was evaluated by the Kalman filtering method, and the error of the subsequent state was estimated optimally, and thus to construct the mathematical model of human behavior. The neural network model was built by wavelet transform function. Moreover, the parameters of human motion feature points extracted before were used as input neurons and input into the neural network for training. Furthermore, the trained neural network was used to quantitatively identify the classification of test sets and thus to determine the index of recognition accuracy. Simulation results show that the proposed method can extract more data and classify more comprehensively than those of traditional methods, so its recognition accuracy is higher and the error rate is lower.