现有人体行为识别算法主要依赖于粗粒度的视频特征，但这些特征不足以有效描述人体行为 的动作构成，从而降低了深度学习模型对易混淆行为的识别能力。该研究提出了一种基于人体部件的 视频行为识别方法，通过学习人体细粒度部件的动作表示，自底向上地学习人体行为视频表征。该方 法主要包含：(1)部件特征增强模块，用于增强基于图像的人体部件特征；(2)部件特征融合模块，用 于融合人体各部件特征以形成人体特征；(3)人体特征增强模块，用于增强视频帧中所有人的人体特 征。该方法在国际标准数据库 UCF101 和 HMDB51 上进行的实验验证结果显示，基于人体部件的视频 行为识别方法与已有方法具有良好的互补性，可以有效提高人体行为识别精度。
High-level action information, such as spatial feature of frames, temporal feature among frames, or human level skeleton features are usually used in existing video action recognition methods. However, these high-level features cannot effectively describe the action composition of human behavior, and thus reduce the ability of deep learning models to recognize confusing behaviors. In this work, video action recognition method based on human body parts is investigated. By learning the action representation of the fine-grained parts of the human body, the video representation of human action was learned from bottom to up level. Specifically, the method mainly includes three modules: (1) body part feature enhancement module, which enhances the image-based human body part feature, (2) body part feature fusion module, which fuses the features of various parts of the human body to form human feature, and (3) body feature enhancement module, which is responsible for enhancing the human body features of all people in the video. The popular datasets of UCF101 and HMDB51 were used for experiments. And the results showed that, the video action recognition method based on human body parts is complementary with current methods, and can effectively improve the accuracy of human action recognition.
XIA Ding, WANG Yali, QIAO Yu. Research of Video Action Recognition Based on Human Body Parts[J]. Journal of Integration Technology,2021,10(5):23-33