基于改进深度学习模型C-GRBM的人体行为识别

Human action recognition based on improved depth learning model C-GRBM

为了提高复杂视频中人体行为识别的性能,在Gate限制玻尔兹曼机(gate restricted boltzmann machine,GRBM)模型基础上提出一种结合卷积神经网络的Convolutional-GRBM(C-GRBM)模型。利用视频图像平稳性的特点,通过不同的卷积核提取可见层不同的特征,提高模型局部特征提取能力,进而得到更好的人体行为识别率;加入池化操作,对卷积层输出的不同位置上的特征进行聚合统计,降低卷积层输出特征量的维度,从而解决原模型参数过多、容易过拟合等缺陷,进而降低人体行为识别复杂度。在人体行为测试库上的测试表明,本文提出的CGRBM模型能够较好地提高人体行为识别性能。

A model of convolutional-gate restricted Bohzmann machine （C-GRBM） combined with a convolutional neural network, which was based on the GRBM model, tion recognition in complicated videos. First, by taking is proposed herein to improve the performance of human acadvantage of the smoothness of video frames, and extracting different visible layer characteristics using different convolution kernels, the ability of extracting local model features and the successful rate of human action recognition can be improved. Second, the operations of pooling were added to obtain the aggregate statistics of features output by the convolntional layer at different positions. In addition, the dimensions of features output by the convolutional layer were reduced to overcome defects such as excessive parame- ters of original model, as well as the high possibility of over-fitting. Consequently, the complexity of human behavior recognition was reduced. Tests based on the human action test database were performed. These tests indicated that the ability of human action recognition can be improved using the proposed C-GRBM model.