人体动作识别在近年来成为许多研究课题的关注点。良好的人体动作识别模型可以应用在生活中的各个领域，例如智能安防，健身设备等日常生活场景中，帮助并优化提升人机交互的体验。为了达到更好识别动作的目标，模型需要适应学习不同的动作特点，应对复杂的环境条件。本文主要研究如何有效的提取运动过程中的人体特征，针对不同的环境影响因素以及识别任务需求，提出合理的解决方案，并将其应用于模型设计中，提高动作识别的性能以及提升用户的交互体验。一、提出基于特征选择模块的图神经网络人体动作识别模型，强化特征提取的过程，使模型整体的识别率和鲁棒性得到提升。二、提出基于密集连接的变换视角人体动作识别模型。模型可以适应更复杂的动作观测环境，以及更有效的利用动作特征信息。三、提出基于波形拟合方法的实时动作评价模型。该模型可根据任务需求，将动作评价结果应用到现实场景中，并有良好的性能表现。

Human action recognition has become the focus of many research topics in recent years. A good human action recognition model can be applied in various areas , such as intelligent security, fitness equipment and other daily life scenarios to help and optimize the experience of human-computer interaction. In order to achieve the goal of better recognition of actions, the model needs to be adapted to learn different action characteristics and deal with complex environmental conditions. This article mainly studies how to effectively extract human body characteristics during actions, propose reasonable solutions for different environmental factors and recognition task, and apply them to model design to improve the performance of action recognition and enhance user interaction experience. The main work and contributions are as follows:A graph-based neural network human action recognition model based on the feature selection module is proposed, which strengthens the process of feature extraction, and improves the overall recognition rate and robustness of the model.The differentiation of human body structure and the differentiation between actions brings many challenges to action recognition. In order to construct the rich features, this model uses the graph neural network as the model backbone, the human body structure topology diagram as the data input of the graph neural network, and uses the dual-stream network structure design to use bone feature information and joint feature information as graph nodes. The feature selection module is introduced to assign the weight of the feature channels, so that the model can learn the global information of the action, and more selectively focus on the human body components with high action participation.A human action recognition model based on dense connections with a view-invariant graph representation method. The model can adapt to a more complex action observation environment, and make more effective use of action feature information.Ignoring the context in the action recognition process can lead to confusing results, and the action information observed from different perspectives is different, resulting in the lack of information. Similar problems also exist in some small movement actions. This model extracts the motion patterns of joints and bones at the input end of the model, and concatenate them with spatial features to represent the spatial and temporal information. Use view-invariant graph representation method to adaptively generate more diverse action features from different perspectives to deal with perspective problems. Introduce dense connections between the graph convolution blocks to enhance the reuse of action features.Real-time action evaluation model based on wave fitting method. The model can apply the action evaluation results to real scenes according to task requirements, and has good performance.In real-life scenarios, affected by equipment hardware, data collection and other factors, it is necessary to ensure that the application has immediate result feedback and a good user experience. Proposes a real-time action evaluation model based on the wave fitting method. The data collected by the front-end gesture recognition framework is parsed into time series waveform, and the action evaluation results are modeled in the spatial-temporal dimension, combined with machine learning methods to train the weights, which can be applied in real-life scenarios.