步态识别作为一种新兴生物识别技术，因其与传统人脸识别和指纹识别等技术相比具有识别距离远、伪装困难等优点，得到了国内外专家学者的广泛关注与研究。但运动目标的衣着、移动速度及与采集设备间的角度发生变化时，会对识别的准确率产生较大影响。为此，论文对基于卷积神经网络(Convolutional Neural Network，CNN)的步态识别方法进行了深入研究，为解决上述技术难题，提出了一种基于虚拟三维卷积神经网络(P3D CNN)的步态识别算法，具有重要的理论研究意义及工程应用价值。 论文首先建立了步态图像的样本库，采用背景减除法对其中的运动目标进行了提取，通过膨胀和腐蚀等形态学处理方法完成了噪声和空洞等干扰的消除，并对步态图像进行了归一化处理。然后，分别对二维卷积神经网络(2D CNN)和三维卷积神经网络(3D CNN)的步态识别模型进行了详细探讨，并对上述模型的卷积核大小、数目及网络层数等超参数进行了设计，实验结果表明，2D CNN的运行速度较快，但由于只采集空间信息，导致识别率较低，而3D CNN因其同时对空间和时间域特征进行提取而提高了识别率，但引起参数数量增加，导致训练时占用内存多；最后，论文对传统3D CNN进行了改进，提出了一种基于P3D CNN的步态识别算法，在保证较高识别率的基础上，降低了计算复杂度，加快了运行速度；同时，为了减轻衣着变化对于识别率的影响，对步态图像进行了背景化处理，并对处理前后的识别效果进行了对比，指出经过背景化处理后可以显著提高识别率。 论文以Tensorflow为开发平台，采用Python编程语言结合C++完成了基于卷积神经网络的步态识别系统的开发，对上述理论分析进行了实验验证。实验结果表明，所提出的P3D CNN步态识别算法既能获得较高的识别率又可以降低占用的内存资源；同时，对步态图像进行背景化处理后，可大大减弱运动目标衣着对识别率的影响。

As an emerging biometric technology, gait recognition has the advantages of long distance recognition and difficult camouflage compared with traditional face recognition and fingerprint recognition technologies. It has received extensive attention and research from experts and scholars at home and abroad.However, when the clothing target, the moving speed, and the angle with the collecting device change, the accuracy of the recognition will be greatly affected.To this end, an in-depth study on the gait recognition method based on Convolutional Neural Network (CNN) was made. To solve the above technical problems, a step based on virtual three-dimensional convolutional neural network (P3D CNN) is proposed. State recognition algorithm has important theoretical research significance and engineering application value.It has important theoretical research significance and engineering application value. Firstly, a sample library of gait image was sestablished, and uses the background subtraction method to extract the moving targets. The morphological processing methods such as expansion and corrosion complete the elimination of noise and holes, and normalize the gait images. Then, the gait recognition models of two-dimensional convolutional neural network (2D CNN) and three-dimensional convolutional neural network (3D CNN) are discussed in detail. The hyperparameters such as the size, number and network layer of the convolution kernel of the above model are designed. The experimental results show that the 2D CNN runs faster, but because only the spatial information is collected, the recognition rate is lower, and the 3D CNN improves the recognition rate because it simultaneously extracts the spatial and temporal domain features, but causes the number of parameters to increase. , resulting in more memory during training. Finally, the paper improves the traditional 3D CNN and proposes a gait recognition algorithm based on P3D CNN. On the basis of ensuring higher recognition rate, the computational complexity is reduced and the running speed is accelerated. At the same time, in order to reduce the influence of clothing change on the recognition rate, the gait image was background-processed, and the recognition effects before and after processing were compared. It was pointed out that the background rate could significantly improve the recognition rate. The thesis uses Tensorflow as the development platform, and uses Python programming language combined with C++ to complete the development of gait recognition system based on convolutional neural network. The above theoretical analysis is experimentally verified. The experimental results show that the proposed P3D CNN gait recognition algorithm can obtain higher recognition rate and reduce occupied memory resources.At the same time, after the background image of the gait image is processed, the influence of the clothing target on the recognition rate can be greatly reduced.