基于深度学习的三维空间的人体行为图像扫描算法研究

Research on scanning algorithm of human behavior image in 3D space based on deep learning

为提高人体行为图像中识别分辨精准度,提出基于深度学习设计的一种三维空间的人体行为图像扫描算法.采用颜色-深度物体识别技术获取三维空间物体上的数据并判别,结合深度卷积神经网络算法处理数据及寻找数据特征,为解决输出的图像在一定程度上存在失真的问题,提出改进的单层卷积神经网络-空间金字塔池化层-递归神经网络算法,即利用空间金字塔池化层代替卷积神经网络中的池化层,直接对图像的卷积特征池化,使得识别系统设计简单并且不需要重构三维目标的三维模型,以完成人体行为图像扫描识别.仿真实验结果表明,算法可有效对三维空间中人体行为图像数据挖掘,数据特征平均分辨率达到93.7%,识别准确率达到了90.8%,具有一定的应用价值.

In order to improve the accuracy of recognition in human behavior image,a scanning algorithm of human behavior image in 3D space based on deep learning is proposed.The color-depth object recognition technology is used to obtain the data from 3D space object,and the depth convolution neural network algorithm is used to process the data and find the data features.In order to solve the problem of output image distortion,this paper proposes an improved single-layer convolution neural network-space pyramid pooling layer-recurrent neural network algorithm,which uses the inner layer of the spatial pyramid pooling layer replace the pooling layer of the convolutional neural network to complete the scanning and recognition of human behavior images.This method directly pools the convolutional features of the image and it makes the design of the recognition system simple and does not need to reconstruct the three-dimensional model of the three-dimensional target.Simulation results show that the proposed algorithm can effectively mine human behavior images in 3D space with 93.7%average resolution and 90.8%recognition accuracy.The research results have practical certain application value.