Feng Tian

In order to inherit and promote intangible cultural heritage, especially Chinese traditional martial arts, this study designs a martial arts action recognition framework, aiming to allow more people to learn about and study this rich cultural tradition more conveniently. The framework adopts a series of advanced computer vision and deep learning algorithms, including YOLOv5s, DeepSORT, AlphaPose, and ST-GCN, which achieve efficient recognition and classification of martial arts actions. Due to ST-GCN's overemphasis on local spatiotemporal dependencies, it might overlook global information. To address this concern, this study integrates the polarized self-attention mechanism into STGCN. This approach distinguishes vital information, preventing the equal treatment of all features and excessive attention to secondary attributes.And we employed the RAdam optimizer to improve the robustness of the proposed model.The improved model can extract skeletal keypoint motion features more effectively, which is used to distinguish similar actions. Experimental results show that our optimized model has better recognition accuracy on the public dataset, indicating that the model can enhance the recognition rate for similar actions.

Accurate scoring is of paramount importance in the martial arts community for training and competitions. Traditional methods rely on manual observation and subjective judgment, which are resource-intensive and prone to errors. To overcome these challenges and provide a more intuitive quantitative representation for scoring martial arts movements, we propose a martial arts recognition and scoring system developed using artificial intelligence technology. The system consists of a codec architecture and two heads that focus on scoring and action classification respectively. To assess the effectiveness of this method, we collected and annotated a dataset of martial arts techniques for training and model evaluation. Experimental results demonstrate the system's high accuracy in martial arts technique scoring and recognition. The system provides a non-invasive, fast, and accurate way to assess the quality of martial arts movements, which has a large number of application scenarios in martial arts education, training analysis, and event fairness.

There are a large number of Chinese Kung Fu enthusiasts worldwide. Offline learning has been the primary method for teaching and learning Chinese Kung Fu due to the inability of online 2D videos to convey spatial information and provide feedback. To address the issues of the high cost of offline teaching, fixed perspectives in 2D videos, lack of spatial path feedback, and the absence of practical Kung Fu teaching systems, this paper proposes Kung Fu Metaverse: a new movement guidance training system. A digital character in the system is given nearly a hundred kinds of Chinese Kung Fu. Learners can observe the static postures and dynamic transitions of martial arts from various angles. Composite navigation module of our system provides learners with a feedback of spatial path teaching method so that they can receive real-time feedback when training. We recruited volunteers and conducted two experiments. The first compared the cognitive experience of volunteers in three different modes, while the second investigated the impact of voice prompts, trajectory settings, arrow indicators, and free gestures on learning outcomes. The results indicate that volunteers demonstrated significantly higher levels of satisfaction, comfort, and interactivity in the Mixed Reality (MR) mode. Compared to Virtual Reality (VR), MR allows free movement and significantly reduces adverse effects such as sickness and dizziness. The addition of the composite navigation module improves the movement precision during practice. This work is a pioneering system that integrates nearly a hundred sets of traditional Chinese Kung Fu routines with a low-cost device.

This paper presents a cognitive psychology experiment to explore the differences between 2D and virtual reality (VR) film editing techniques. We recruited sixteen volunteers to view a range of different display modes and edit types of experimental material. An electroencephalogram (EEG) was recorded simultaneously while the participants watched. Subjective results showed that the VR mode reflects higher load scores, particularly in the effort dimension. Different editing types have no effect on subjective immersion scores. The VR mode elicited stronger EEG energy, with differences concentrated in the occipital, parietal, and central regions. On the basis of this, visual evoked potential (VEP) analyses were conducted, and the results indicated that VR mode triggered greater spatial attention, while editing in 2D mode induced stronger semantic updating and active understanding. Furthermore, we found that while the effect of different edit types in both display modes is similar, cross-axis editing triggered greater cognitive violations than continuity editing, which could serve as scientific theoretical support for the development of future VR film editing techniques.

Many researchers have measured the differences in electroencephalography (EEG) while viewing 2D and 3D movies to uncover the neuromechanism underlying distinct viewing experiences. Using whole-brain network analyses of scalp EEG, our previous study reported that beta and gamma bands presented higher global efficiencies while viewing 3D movies. However, scalp EEG is influenced by volume conduction, not allowing inference from a neuroanatomy perspective; thus, source reconstruction techniques are recommended. This paper is the first to measure the differences in the frontal-occipital networks in EEG source space during 2D and 3D movie viewing. EEG recordings from 40 subjects were performed during 2D and 3D movie viewing. We constructed frontal-occipital networks of alpha, beta, and gamma bands in EEG source space and analyzed network efficiencies. We found that the beta band exhibited higher global efficiency in 3D movie viewing than in 2D movie viewing; however, the alpha global efficiency was not statistically significant. In addition, a support vector machine (SVM) classifier, taking functional connectivities as classification features, was built to identify whether the frontal-occipital networks contain patterns that could distinguish 2D and 3D movie viewing. Using the 6 most important functional connectivity features of the beta band, we obtained the best accuracy of 0.933. Our findings shed light on uncovering the neuromechanism underlying distinct experiences while viewing 2D and 3D movies.

The intensity of feeling and comfort are two contrasting aspects of 3D movies. This paper studies 3D comfort according to crosstalk, basic parallax, space motion, and depth of field based on pixels. A method for calculating 3D motion based on the XYZ axes is presented, and a pixel-based evaluation model of 3D comfort is established with parallax, motion, and other indexes. In order to support the production and evaluation process of 3D films, a 3D comfort evaluation tool is designed according to time series. As demonstrated experimentally, basic parallax, depth of field, and average crosstalk have certain impact on the entire 3D comfort, while space motion has a relatively salient effect. When applying our method, a steady control on 3D comfort is found in high-quality 3D films: About 92% of comfort is within the range of 1-3.5 points for Transformers: Age of Extinction (USA), whereas about 74% of comfort is within the range of 1-3.5 points for The Monkey King (China). Large depth of field for slow moving lens and small depth of field for fast moving lens were used in Avatar and Transformers, and these 3D art rules are not followed closely in The Monkey King. As a basic tool for 3D film production, our proposed method can be applied in creating, shooting, producing, and monitoring for a 3D film.