基于轻量化卷积神经网络的人体动作识别

Human activity recognition based on lightweight convolutional neural network

针对传统卷积神经网络模型LeNet识别准确率低,占用内存大等问题,提出了一种基于轻量化卷积神经网络的人体动作识别模型(human activity recognition net,HARNet)。首先,利用MobileNetV2模型参数量和计算量小的特点,利用迁移学习方法,将预训练好的权重参数迁移到MobileNetV2模型中,最后添加全连接层构建了HARNet,实现了对日常行为动作的准确识别和分类。实验结果表明,该模型动作识别平均准确率可达89%,相比于传统卷积神经网络LeNet,准确率更高,且训练好的模型内存大小仅8.97 MB,验证了该模型的有效性。

For the problems of low recognition accuracy and large memory consumption of the traditional convolutional neural network model LeNet,a human activity recognition net(HARNet)based on lightweight convolutional neural network was proposed.Firstly,the features of the MobileNetV2 model with a small number of parameters and calculations were used;then the transfer learning method was applied to transfer the pre-trained weight parameters to the MobileNetV2 model.Finally,a fully connected layer was added to build a HARNet,which achieved more accurate recognition and classification of daily behavioral actions.The experimental results show that the average accuracy of activity recognition of the model can reach 89%,which is higher than the traditional convolutional neural network LeNet,and the memory size of the trained model is only 8.97 MB,which verifies the effectiveness of the model.