Department of Physical Education and Research of Lanzhou University, Lanzhou, Gansu, China.
Comput Intell Neurosci. 2023 Jan 25;2023:6116144. doi: 10.1155/2023/6116144. eCollection 2023.
Artificial intelligence and deep learning have attracted much attention from researchers in industry and academia. The volleyball movement standardization and recognition model involve the application of artificial intelligence and deep learning. In order to solve the problem that human action in volleyball video is continuous and effective spatial and temporal features need to be extracted from the video stream, the Inception module is decoupled and heterogeneous, replacing the original 5 × 5 convolutional structures with two 3 × 3 convolutional structures, as well as replacing the 3 × 3 convolutional structures with 1 × 3 and a 3 × 1 convolutional structure with internal parameter optimization to ensure the accuracy of recognition. The model uses the input motion video RGB map as the spatial input and the optical flow map as the temporal input, and the two are weighted 1 : 1 for feature fusion. Experiments are conducted on the volleyball action video and homemade dataset in UCF101, and the experimental data show that the accuracy of the DNet volleyball action standardization recognition model proposed in this paper is 94.12%, which proves that the method improves the recognition ability of the model while speeding up the training speed. The research presented in this paper provides important theoretical guidance for artificial intelligence and deep learning.
人工智能和深度学习引起了工业界和学术界研究人员的广泛关注。排球动作标准化和识别模型涉及人工智能和深度学习的应用。为了解决排球视频中人类动作是连续的问题，需要从视频流中提取连续有效的时空特征，该模型将 Inception 模块解耦和异构化，用两个 3×3 卷积结构替换原始的 5×5 卷积结构，同时用 1×3 和 3×1 卷积结构替换 3×3 卷积结构，并进行内部参数优化，以保证识别的准确性。模型使用输入运动视频的 RGB 图作为空间输入，光流图作为时间输入，并对两者进行 1∶1 的加权特征融合。在 UCF101 的排球动作视频和自制数据集上进行实验，实验数据表明，本文提出的 DNet 排球动作标准化识别模型的准确率为 94.12%，证明该方法在提高模型识别能力的同时加快了训练速度。本文的研究为人工智能和深度学习提供了重要的理论指导。
