基于改进残差网络的儿童动作分类

Action Classification of Children Based on Improved Residual Networks

对室内无人或室内弱监督情况下的儿童动作分类可以很好的预防儿童危险行为的发生。传统姿态分类公开数据集KTH,HMDB中人物背景单一。对此,采用实景拍摄的A2D公开数据集,人物背景较为复杂。传统残差网络在此数据集的中提取有效特征的能力较差。对此,首先将数据集进行多种数据增强,在残差网络Resnet34中引入通道与空间上的注意力机制CBAM(Convolutional Block Attention Module),最后将提取特征送入全连接层,实现对图片中的儿童动作的分类。结果显示,改进残差网络在测试集上的准确率为79.8%,比传统残差网络的准确率提升了10%,满足室内儿童动作分类的要求。

Classifying children’s movements when there is no indoor or weak indoor supervision can well prevent the occurrence of children’s risky behaviors. The traditional pose classification public dataset KTH,HMDB has a single background of characters. In this regard, the A2D public dataset used in real action shooting has a complex background of the characters. Traditional residual networks have a poor ability to extract valid features from this dataset. In this regard, a variety of data enhancements are made to the dataset, and the attention mechanism CBAM(Convolutional Block Attention Module) on the channel and space is introduced in the residual network Resnet 34, and finally the extracted features are sent to the fully connected layer to realize the classification of children’s actions in pictures. The results show that the accuracy of the improved residual network on the test set is 79.8%, which is 10% higher than that of the traditional residual network, which meets the requirements of indoor children’s motor classification.