Various RNN-based HAR studies.

Human Action Recognition (HAR) is a pivotal area in computer vision, video surveillance, and human-computer interaction (HCI), driven by the need for efficient and accurate models to enhance HCI experiences. Traditional HAR methods often rely on hand-crafted features and shallow learning techniques, which limits their ability to capture complex patterns. In contrast, this study proposes an efficient HAR model that leverages deep neural networks, specifically a combination of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), to enhance HCI through AI-powered action understanding. The model employs a pre-trained EfficientNetB7 network to extract rich spatial features from video frames, followed by a Long Short-Term Memory (LSTM) network to capture long-range temporal dependencies. This architecture enhances recognition accuracy while reducing computational complexity, making it highly suitable for HCI applications. Experimental results demonstrate the superior performance of the model, achieving a classification accuracy of 97.8% on the UCF101 dataset and 80.1% on the HMDB51 dataset, outperforming state-of-the-art HAR models. The proposed model eliminates the need for auxiliary assistive techniques like data augmentation, highlighting its efficiency and tremendous potential for real-world HCI applications that rely on accurate and efficient recognition of human actions.人体动作识别（Human Action Recognition，HAR）是计算机视觉、视频监控与人机交互（Human-Computer Interaction，HCI）领域的核心研究方向，其发展源于对高效精准模型的迫切需求，旨在优化人机交互体验。传统人体动作识别方法多依赖手工设计特征与浅层学习技术，极大限制了其捕捉复杂动作模式的能力。与之相对，本研究提出一款高效的人体动作识别模型，该模型依托深度学习网络架构——具体为卷积神经网络（Convolutional Neural Networks，CNNs）与循环神经网络（Recurrent Neural Networks，RNNs）的组合——通过AI驱动的动作理解能力优化人机交互体验。该模型采用预训练的EfficientNetB7网络从视频帧中提取丰富的空间特征，随后接入长短期记忆（Long Short-Term Memory，LSTM）网络以捕捉长程时间依赖关系。该架构在提升识别精度的同时降低了计算复杂度，使其高度适配人机交互应用场景。实验结果验证了该模型的优异性能：在UCF101数据集上的分类准确率达97.8%，在HMDB51数据集上达80.1%，其性能优于当前主流的人体动作识别模型。所提模型无需借助数据增强等辅助技术，凸显了其高效性与巨大应用潜力，可适配依赖精准高效人体动作识别的真实人机交互场景。