基于优化卷积神经网络结构的人体行为识别

为了提高卷积神经网络对非线性特征以及复杂图像隐含的抽象特征提取能力,提出优化卷积神经网络结构的人体行为识别方法。通过优化卷积神经网络模型,构建嵌套Maxout多层感知器层的网络结构,增强卷积神经网络的卷积层对前景目标特征提取能力。通过嵌套Maxout多层感知器层网络结构可以线性地组合特征图并选择最有效特征信息,获取的特征图经过矢量化处理,分类器Softmax完成人体行为识别。仿真实验结果表明,该方法对人体行为识别准确率取得较好结果。

HUMAN BEHAVIOR RECOGNITION BASED ON OPTIMIZED CONVOLUTIONAL NEURAL NETWORK STRUCTURE

To improve the abstract feature extraction ability of convolutional neural networks on nonlinear features and complex images, a human behavior recognition method based on the structure of convolutional neural networks is proposed. By optimizing the convolutional neural network model, the method constructed the network structure of the nested Maxout multi-layer perceptron layer, enhanced the convolutional neural network’s convolutional layer to extract the foreground target features, and nests the Maxout multi-layer perceptron layer network structure. The feature map could be linearly combined to select the most effective feature information, and the acquired feature map was subjected to vectorization processing, and the classifier softmax was used for classification and recognition of human behavior. The simulation results show that the method has a good result on the accuracy of human behavior recognition.