AI Networking Cookbook: Practical recipes for AI-assisted network automation and development
TL;DR
This study develops a skeleton-imposed images-based CNN model, contrasting traditional CNN approaches that typically rely solely on body shape or key point representations, that enhances the accuracy of yoga pose classification, providing a more nuanced understanding of complex poses.
An integrated multi-layer hybrid framework for accurate, real-time posture assessment in healthcare and rehabilitation contexts is proposed, although all solutions trade off accuracy, computational cost, and practical generalizability.
Yoga represents an age-old practice that is beneficial for both psychological and physical health. The yoga promotes self-learning and improper poses can seriously harm muscles and ligaments. The accurate recognition of yoga poses from images remains challenging due to high intra-class similarity, background variations, and redundant feature generated by deep learning models. This paper proposed a model that combined the MobileNet with Hunter-Prey Optimization, and Long Short-Term Memory to detect the efficient and accurate yoga poses. In the proposed model, Mobile Net is used as a lightweight backbone to extract high-level unique and spatial features from yoga images. The extracted feature vectors are then optimized using the HPO algorithm that selects the most relevant and discriminative features and remove redundant feature. The optimized feature is fed into an LSTM model. The LSTM model operates on static image inputs as a gated feature refinement and classification rather than modelling temporal pose transitions. The performance of model is measured on a 6-class subset of the Yoga-82 dataset that allows the controlled experimental parameters. The final results shows that the proposed model obtained an accuracy score of 97.97% performed well as baseline models such as MobileNet and MobileNet + LSTM. The HPO-based feature selection reduces the feature dimensionality up to 69.53% to improved computational efficiency and reduced inference time. The result shows when combined the feature optimization with lightweight MobileNet model then enhances classification performance and maintain the efficiency of model. The proposed model is applicable for real-time yoga pose recognition with future work focused on extending the model to full-scale datasets and video-based temporal modelling.
This study aims to develop a human body posture classification model based on digital images using the InceptionV3 architecture. The dataset used in this study is the MPII Human Pose Dataset, which contains a wide variety of human activities and body postures. The research began with label extraction from a metadata file in .mat format, followed by the selection of the 20 activity classes with the largest number of samples. Subsequently, the dataset was balanced using an undersampling technique, resulting in 140 images for each class. The images were then resized to 224×224 pixels, normalized using the preprocess_input function, and enhanced through data augmentation applied to the training set. The model was developed using a transfer learning approach with InceptionV3 as the base model. Additional layers, including Global Average Pooling, Dropout, and Dense layers, were added to perform multi-class classification. The experimental results showed that the proposed model achieved a test accuracy of 0.8893 with a test loss of 0.5587. Furthermore, the macro-average metrics obtained from the classification report were 0.9018 for precision, 0.8893 for recall, and 0.8877 for F1-score. These results indicate that the model was able to classify most activity classes effectively. However, several classes with similar visual characteristics still caused misclassification, indicating opportunities for further improvement in human posture recognition performance.
With the development of artificial intelligence and deep learning technology, human pose estimation has been widely applied in fields such as medical rehabilitation and motion analysis. To meet the needs of objective assessment of lower limb function in patients with knee joint diseases, we propose a method for detecting and assessing the function of human lower limb keypoints by integrating YOLOv8 with an improved HRNet. Firstly, the YOLOv8s model is used to locate the human region in the image. Then, the HRNet-W32 model, enhanced with a Gating Unit and Keypoint Attention Unit, is employed to detect six lower limb keypoints: left/right hip, left/right knee, and left/right ankle. Finally, based on the geometric relationship of hip-knee-ankle, the flexion-extension angle of the knee joint is calculated, and the sitting-to-standing action is recognized through a threshold state machine. Experimental results show that the improved HRNet achieves a 3.9% and 1.3% increase in mAP@0.5 and PCK@0.3 metrics, respectively, compared to the original HRNet-W32. In 200 sets of 30-Second Sit-to-Stand Test videos, the system achieves an average counting accuracy of 97.84% and an average absolute error of 0.18 times. The correlation between the visual algorithm output and clinical functional status was further evaluated using established clinical indicators, including the Knee injury and Osteoarthritis Outcome Score (KOOS), Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC), Timed Up and Go (TUG) test, and the 30-Second Sit-to-Stand Test.
Human posture recognition is important in sports training because it allows proper evaluation of the technique of an athlete, enhances performance, and reduces the risk of injury. Conventional techniques, including coach manual evaluation and motion capture systems, are generally timeconsuming, expensive, and not scalable. To solve these issues, the suggested approach combines a Random Forest (RF) model with Deep Learning (DL) methods to develop an adaptive and efficient posture recognition system. The hybrid model applies RF for static posture classification and Bidirectional Long Short-Term Memory (BiLSTM) for Modeling sequential dependencies in synthetic sequences of static postures, created by data augmentation. The uniqueness of this method is that RF and DL are combined to improve posture recognition. Improvement in RF was achieved by applying feature selection techniques (RFE) and hybrid tree pruning, reducing computational overhead while improving classification accuracy. These optimizations make the system more efficient and scalable, suitable for large-scale training environments. The RF model’s accuracy improved by 4.7%, from 92.5% to 97.2%, after the application of feature selection and pruning techniques. The outcomes testify to the proficiency of the model with a 99.08% classification accuracy, 99.28% average precision, 99.18% recall, and F1-score of 99.38%. The model also evidenced good performance over a range of sports postures, with maximum accuracy of 99.66% in sections like Axe Throwing. This method is of great benefit compared to conventional techniques by providing a computationally cost-effective, scalable, and highly accurate posture recognition system, which ultimately leads to improved sports training results and the minimization of injury risks.
The proposed Media Pipe–CNN framework provides an efficient, accurate, and marker less solution for automated ergonomic risk assessment, supporting intelligent occupational safety management, continuous workplace monitoring, and the implementation of smart manufacturing systems aligned with Industry 4.0 initiatives.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.