Due to the diversity of human postures and the changes in actions under different perspectives, traditional recognition methods often fail to achieve ideal recognition results. This paper first builds a human posture estimation model based on Res2Net. The pre-trained Res2Net network is used to extract image features, which are further processed by CNN to predict the key posture points of the human body. This paper also uses data enhancement technology to perform random cropping, rotation, and flipping operations on the data to simulate posture changes under different perspectives. Subsequently, an attention guidance mechanism is introduced to optimize the action recognition process. In the action recognition stage, the attention mechanism is used to weight the posture estimation results and extract more effective action features. Finally, the action features are classified through the fully connected layer to identify the action categories of the human body. By combining Res2Net with the attention guidance mechanism, this method is superior to the traditional method based on manual features in posture estimation accuracy, reasoning time and action recognition accuracy. The average posture estimation accuracy of the proposed method reaches 98.036%, and the inference time is shortened. In terms of action recognition accuracy, the proposed method achieves higher accuracy points in 25 groups of scenes, providing new ideas and technical support for research in related fields.