基于卷积神经网络的人体动作识别研究

Human Activity Recognition Using Convolutional Neural Network

卷积神经网络通过自主学习实现特征提取和分类,同时使用了局部连接和权值共享降低网络复杂度并提升学习效率.为了提高了人体动作识别的准确率,采用了卷积神经网络来实现人体动作识别,并在公开数据集中验证此算法的准确率.实验结果表明,人体动作图像分辨率、卷积核和神经网络参数选取对识别准确率有影响;在一般情况下,识别准确率可达99%以上,证实基于卷积神经网络的人体动作识别是一种稳定可行的识别方法.

Convolutional neural network realizes feature extraction and classification through autonomous learning,and uses local connection and weight sharing to reduce network complexity and improve learning efficiency.The convolutional neural network is used to improve the accuracy of human activity recognition and the accuracy of this algorithm is evaluated in this paper.Experimental results show that image resolution of human activity,convolution kernel and neural network parameters have an impact on the recognition accuracy.The recognition accuracy can reach more than 99%,which proves that human motion recognition based on convolutional neural network is a stable and feasible recognition method.