Results of statistical significance test.

With the development of smart wearable devices and deep learning (DL) technology, the monitoring and analysis of daily sports activities of teenagers face new opportunities. At present, traditional CNN (Convolutional Neural Network) models are mostly used for recognition in daily sports activities. It is difficult to capture the temporal relationship between action sequences, and the ability to express important features is weak, resulting in poor recognition accuracy. This paper took badminton as the object, based on the VGG16 (Visual Geometry Group 16) model, and adopted the advantages of the bidirectional learning time series information of the BiLSTM (Bidirectional Long Short-Term Memory) model and the channel and regional feature representation of the CBAM (Convolutional Block Attention Module) module to verify and apply the recognition of badminton movements in daily sports for teenagers. The study first built and optimized the baseline model VGG16, removed the last three fully connected layers, and used VGG16 to extract the deep features of each frame of video image and output feature maps. The CBAM module was then embedded after the last convolutional layer of the VGG16 network, and the feature maps optimized by CBAM were flattened into a time series input vector. Finally, the BiLSTM model is introduced, and the CBAM and BiLSTM are connected in a cascade manner to capture the information of the previous and next dependencies in the video frame sequence and output the action classification results of badminton. The experiment is based on the badminton training dataset in the public dataset Roboflow to explore the action recognition performance in badminton in daily sports activities of teenagers. Experimental results show that the recognition accuracy of the VGG16-BiLSTM-CBAM model has reached 0.98, which is 0.08 higher than the benchmark model VGG16, and F1 has reached 0.96. Experimental results show that combined with the DL model VGG19 and the sequential model BiLSTM, the attention CBAM module can significantly improve the performance of action recognition in youth badminton, promote the safe conduct of sports activities, and provide a good reference for incorrect postures.

随着智能可穿戴设备与深度学习（Deep Learning，DL）技术的发展，青少年日常体育活动的监测与分析迎来了新的机遇。当前，日常体育活动中的动作识别多采用传统卷积神经网络（Convolutional Neural Network，CNN）模型，但这类模型难以捕捉动作序列的时序关联，且重要特征的表达能力较弱，导致识别精度不佳。本文以羽毛球运动为研究对象，以VGG16（Visual Geometry Group 16）模型为基础，结合双向长短期记忆网络（Bidirectional Long Short-Term Memory，BiLSTM）模型对时序信息的双向学习能力，以及卷积块注意力模块（Convolutional Block Attention Module，CBAM）的通道与区域特征表达优势，开展青少年日常羽毛球动作识别的验证与应用研究。本研究首先构建并优化基线模型VGG16：移除其最后三层全连接层，利用VGG16提取视频图像每一帧的深层特征并输出特征图；随后在VGG16网络的最后一个卷积层后嵌入CBAM模块，将经CBAM优化后的特征图展平为时序输入向量；最后引入BiLSTM模型，通过级联方式连接CBAM与BiLSTM，以捕捉视频帧序列中前后依赖的关联信息，最终输出羽毛球动作的分类结果。本实验依托公开数据集Roboflow中的羽毛球训练数据集，探究青少年日常体育活动中的羽毛球动作识别性能。实验结果表明，VGG16-BiLSTM-CBAM模型的识别精度达0.98，较基准模型VGG16提升0.08，F1值达0.96。实验结果证实，结合深度学习（Deep Learning，DL）模型VGG19与时序模型BiLSTM的注意力CBAM模块，可显著提升青少年羽毛球动作识别性能，助力体育活动安全开展，并为纠正错误动作姿态提供良好参考。