An in-depth learning-based approach is designed to develop the ability to recognize human behavior on the move. We introduce 3D residual structures and create 3D residual models. In order to get the most out of the data relationship of several consecutive frames, this study introduces 3D techniques for assigning different values to the existing frames. Experiments show that both structures improve recognition performance. For the 3D residual model, 3D attention model, and 3D attention residual model, this study proposes two model fusion strategies: average and weighted. Among them, the weighted fusion is to give a higher fusion proportion to the high accuracy model by using the model weight calculation method designed in this study. The experimental results show that the additive fusion strategy based on feature contribution has an obvious improvement effect on the test results of the two benchmark datasets, with an increase of more than 2 points, including an increase of 2.69 on HMDB51. The effect of splicing and fusion strategy has also increased by more than 1 point, including 1.34 on UCF101 dataset and about 1.9 on HMDB51. It is proven that deep learning can effectively recognize human behavior in sports. ? 2022 Xiwei Liu.
展开▼
机译：一种基于学习的深入方法旨在培养识别移动中人类行为的能力。我们引入了 3D 残差结构并创建了 3D 残差模型。为了充分利用多个连续帧的数据关系，本研究引入了为现有帧分配不同值的 3D 技术。实验表明，这两种结构都能提高识别性能。针对三维残差模型、三维注意力模型和三维注意力残差模型，提出了平均和加权两种模型融合策略。其中，加权融合是利用本研究设计的模型权重计算方法，为高精度模型赋予更高的融合比例。实验结果表明，基于特征贡献的加法融合策略对两个基准数据集的测试结果有明显的提升效果，提升幅度超过2个百分点，其中HMDB51提升了2.69%。剪接融合策略的效果也提升了1个百分点以上，其中UCF101数据集提升了1.34%，HMDB51提升了约1.9%。事实证明，深度学习可以有效地识别人类在运动中的行为。?2022 刘希伟.
展开▼
