Human Pose Recognition Using Deep Learning

The recognizing human body posture is an enduring subject in the computer vision-based deep learning field that has highlighted numerous challenges in the past. Human activity analysis is useful in numerous fields, including security, video games, physical treatment, and so on. Exercises or yoga postures are one of the difficulties in estimating human poses. The current study designates a hybrid deep learning methodology for properly recognizing various yoga poses. A publicly available dataset of 27 individuals captured in an in-house premises, serene of 19 females and 8 males, performing each of the ten yoga poses is used. Yoga poses, such as “garland pose”, “happy baby pose”, “head to knee pose”, “lunge pose”, “mountain pose”, “raise arm pose”, “seated forward bend”, “staff pose”, “plank pose”, and “standing forward bend” are used. A hybrid deep learning model based on convolutional neural network (CNN) and long short-term memory (LSTM) is giving some promising results for yoga pose recognition on real-time videos. Using CNN extracting features from key points of each frame obtained from OpenPose and LSTM provides temporal predictions. The system achieves a test accuracy of 97.38% after aggregating predictions over 45 frames, indicating its robustness in dealing with temporal variations. Real-time testing with 100 people offers an accuracy of 97.22%, demonstrating the model’s practical application in scenarios such as yoga training and health monitoring.