Asana Aesthetics
Abstract - We all know how impactful a good yoga session can be. The trend these days is that people are preferring online yoga sessions more and more after the pandemic. But technically, these sessions are not immersive like live yoga sessions. So, to cover the gap, we are trying to build a machine learning model which will assist the users while they practice yoga in the online mode. Our model will track the user's yoga footage and will analyze various aspects of the asana and it will give them some feedback regarding what changes can be made to perfect their pose.
Key Words: Machine Learning, Deep Learning, Computer Vision, Yoga Asana Detection, Pose Correction.
1. INTRODUCTION
In conventional yoga, they say that the two instruments needed are the body and the mind. But in online yoga, an additional instrument is also essential, which is the computer. Ideally, the computer should have minimal impact on the learning process of the learners, and the yoga teachers should feel empowered in the online mode andnothandicapped.Ifwemustcounterthisshortcoming, then we need a robust mechanism of hardware and software which will not only enhance connectivity between student and the teacher, but also assist them enhance their learning experience. It would be even good if the computer has the capability to correct the learner from time to time by giving them feedback. This is where theroleofmachinelearninganddeeplearningcomesinto thelimelight.Eitherthereal-timeorpre-recordedfootage is fed into a machine learning model, which will then extract the key features from the footage, then analyze them,classifythem,andfinallypointoutthesimilaritiesor the anomalies to the user so that they can make changes. Forthis,avarietyofmachinelearningmodelsareutilized, including support vector matrices (SVM), convolutional neural networks (CNN), and regression analysis. Also, other technologies like OpenCV can also be used to support the system. Now, let us see some of the preexistingmodelsthatweperceivedaspartofourliterature survey.
2. LITERATURE SURVEY
Sinha [1] proposed Deep learning-based techniques for identifying inappropriate yoga postures. With this approach,usersgettochoosethepracticeposetheywant toperformanduploadvideosofthemselvesperformingit. Theuserposeisgiventomodelsthatare beingtrained to outputany unusual angles between theuser pose andthe actual pose. By highlighting the yoga pose's shortcomings usingtheobtainedoutputs,theapplicationcanguideusers onhowtocorrectthepose.PredictionAccuracyis99.58%. ThekeypointsareextractedusingMulti-LayerPerceptron andKerasmulti-personposeestimation.
Varsha [2] used KNN classifiers and "PoseNet". An individualcanlearnthebestpossiblemethodtocarryout theparticularyogaposethey'retryingtoachievebyusing such deep learning algorithms. Human pose estimation is used to analyze a person's yoga posture using computer vision algorithms and "Open Pose". The proposed system distinguishes between the actual and target positions and accurately corrects the user by providing real-time visual output and the necessary instructions to change the identifiedpose.PredictionAccuracyis99.51%.
Santhosh [3] applied Deep learning algorithms to effectively detect and distinguish different yoga positions. The 85 videos make up the selected dataset each have 15 persons performing six yoga asanas. The user's keypoints are primarily retrieved using the Mediapipe library. A "LSTM '' and "CNN" combination has been used as a deep learning model to identify yoga asanas using real-time observed footage. The “CNN” layer detects the key points fromwhichthefeaturesareobtained,andafterthatLSTM is used, which recognises when a series of frames occurs so that predictions may be generated. The poses are then categorizedasbeingcorrect orimproper.Thesystemwill provide text/speech feedback to the user if the correct poseisrecognized.PredictionAccuracyis99.53%.
Rutuja [4] employed a technique that begins by running theinputimagesvia"CNN"classifierthathasbeentrained to detect humans. After recognizing yoga poses, the pose estimation network looks for trained key points. The image can then be displayed to the user by the computer using pointers that identify different parts of the user body. The OpenPose library is used in this case to extract
the key points of the pose from the video frame. Yoga poses are detected and corrected using PostNet. Machine learning prediction is used by SVM classifiers to enhance ML algorithms' performance. In order to find similarities between the user's postures and the human skeleton poses, CNN compares the two poses. The prediction accuracyis99%.
Abhishek [5] developed a pose recognition system, which assists and corrects users while performing Yoga. A datasetcalled"YOGI"wascreated,thiscomprised10yoga postures,eachofwhichhadbetween400and900images. Alongwiththat,anotherdatasetforYogaMudradetection wasalsodeveloped. Thatdatasetcontains5 mudras,with around 500 images for each mudra. Two different processeswerecarriedout.Firstonewasfordetectingthe skeleton structure while performing the yoga pose, and the other was for detecting the hand mudras. Later different angles between the joints extracted from the skeleton were measured and then the yoga pose was classified.Anumberofdeeplearningmodelswereusedin theprocessandamongallthetestedmodels,XGBoostwith Random Search cross validation was the most accurate model,givinganaccuracyof99.2%.
Pranjal [6] used the Audiopipe Library. It is used to look over the postures of Surya Namaskar. Advanced software isusedtodetectthestandinginrealtime.ThesePostures areclassifiedintodifferentgroupsofSuryaNamaskar.The form is recognized by the classification divider as being one of the following: ‘Ashtanga Namaskar’, ‘Dandasana, Bhujangasana’, and ‘Svanasana’, as well as ‘Pranamasana’, ‘Hasta Padasana’, ‘Hasta Uttanasana’, ‘Ashwa-Sanchalan Asana’. The correct yoga pose (Surya Namaskar) is identifiedusingthismodel,whichwascreatedusingdeep learning-based approaches (CNN). The model has an accuracy score of 0.75 and a prediction accuracy of 98.68 percent.
Josvin [7] used transfer learning using the VGG16 architecture. To enhance the accuracy, he trained ImageNet weights in conjunction with a DNN classifier. Convolutional neural networks (CNN) and transfer learning, two deep learning approaches, have been employedtodevelopasystemcapableofdetectingayoga pose from an image or frame of a video. The model was trained using photos of 10 different asanas, and the predictionaccuracywasassessed.Withanaccuracyrateof 82%,thepredictionmodelsupportedbytransferlearning displaysencouragingresults.
Satyam and Animesh [8] used the Tensorflow MoveNet Thunder model. The MoveNet Thunder model is an ultrafast and accurate model that estimates the real-time pose and detects it, thereby allowing them to correct it. MoveNet is a model for identifying poses that uses 17
critical sites on the human body to distinguish poses. These key points are then converted to vectors and then "cosine rule" is used to estimate the error in the pose as compared to the ideal pose. The model showed a promising85%predictionaccuracy.
Utkarsh and Dr. Shikha [9] used MediaPipe to stream either live feed or recorded feed into OpenCV where 33 Keypointsinthehumanbodywereextracted.Laterthese Key points were passed through various classificationbased Machine learning algorithms, and eventually, Logistic Regression Classifier produced the most optimizedresultswith94%Predictionaccuracy.
Ajay[10]usedahumanjointslocalizationmodeltolocate all the important joints in the human body while examining the feed. Later a "CNN" model was used for classifyingthese jointsinto different asanas.Based on the classification, they displayed visual and textual feedback to the users, so as to make them aware of the shortcomingsinthewaytheywereperformingyoga.This modelgavearespectable95%classificationaccuracy.
Deepak and Anurag [11] used an instantaneous human pose estimation library called OpenPose to distinguish different joint areas either in the live feed or in recorded feed. Once the joint extraction was done, the output was then fed into a "CNN" plus "Long Short-Term Memory" model, which classifies the pose into a specific asana and givesdecisivefeedbackaswell.TheCNNplusLSTMmodel gaveagood98.58%PredictionAccuracy.
Shruti[12]usedaninstantaneous humanposeestimation library called OpenPose to extract essential key points fromeitherpre-recordedorreal-timefeed.Thentheyused an amalgamated model consisting of "CNN" classifier and "LSTM" model to classify the pose into a particular asana based on the closeness of the extracted key points of the current pose to those of the ideal pose/asana that the model was pre-trained with. This hybrid model produced a99.38%predictionaccuracy.
Ze Wu [13] suggested a full-body posture modeling and quantitativeevaluationapproachtoidentifyandrateyoga poses. With the use of 11 IMUs fixed on the body, quaternion format data has been measured for human posture using a wearable device. As the first classifier for theclassificationofyogaposes,BackPropagationArtificial Neural Networks (BP-ANN) were selected. To assign postures to a category, flexible fuzzy partitioning was performed using FCM as the second classifier. They carried out both the identification of data frames and instances of posture. In the data frame recognition test, 30% of the data were randomly selected from the databasetotraintheBP-ANNandFCMclassifiers,andthe remaining 70% of the data had a recognition accuracy of 89.34%.
Gochoo [14] proposed an innovative yoga posture detectionbasedonInternetofThings(IoT)using“DCNN”s as well as wireless sensors with low-resolution infrared sensors. They achieved a breakthrough experimental outcome by using a low-resolution infrared sensor based WSN for the initial time to recognise up to 26 different yoga poses. The hardware for the suggested system is inexpensive and small. The suggested system is a cuttingedge posture recognition system that is device-free, portable,inexpensive,dependable,andaccurate.
Usama's [15] objective is to identify yoga positions performed by people using Microsoft Kinect and to compare such poses to the actual data. They have established a model of real data that the postures have been discovered using for a certain set of poses. Their main challenge was detecting joint coordinates. Through the use of Kinect capture, the poses are detected. With over97%accuracy,theywereabletoidentifyallanglesin betweendifferentbodysegmentsinyogapostures.
Waseem [16] proposed a powerful classifier system that can identify yoga postures using an RGB camera. The recognition system first makes use of the Blaze Pose frameworkinordertopickoutsignificantpointsfromthe input stream. In order to produce key points that are independent of frame and resolution, the obtained key points are subsequently converted using feature extraction techniques. A unique DL model of "CNN" plus "LSTM"isthenusedwiththeprocessedkeypoints.Finally, theyhaveachieved98.65%predictionaccuracy.
Carreira [17] developed the "Iterative Error Feedback" framework and is interested in using feedback to build predictorsthathaveanaturalabilitytodealwithcomplex, structured output spaces. By training stratified feature extractorsoverasharedarea,theywereabletodevelopa generalmodeltorepresentaformidablestructureinboth inputandoutputspaces.Theydidthisbyintroducingtopdown feedback, which predicts what is wrong with their current estimate and corrects it iteratively rather than trying to forecast the goal outputs directly as in feedforwardprocessing.
3. CONCLUSIONS
The purpose of this review was to get a better understanding of the trends in Yoga Pose Detection technology.Ifwelookbackovertheyears,wecanseehow varied approaches were taken to carry out the process. This knowledge will help us understand the various options that lay in front of us, and then plan a project roadmapforthisendeavor.
REFERENCES
[1] Vivek Anand Thoutam, Anugrah Srivastava, Tapas Badal, Vipul Kumar Mishra, G. R. Sinha, Aditi Sakalle, HarshitBhardwaj,and ManishRaj:"Yoga PoseEstimation andFeedbackGenerationUsingDeepLearning",2022
[2] Varsha Bhosale, Pranjal Nandeshwar, Abhishek Bale, Janmesh Sankhe: "Yoga Pose Detection and Correction usingPosenetandKNN",2022
[3] Debabratha Swain, Santhosh Satapathy, Pramoda Patro, Aditya Kumar Sahu: "Yoga Pose Monitoring System usingDeepLearning",2022
[4] Rutuja Gajbhiye, Snehal Jarag, Pooja Gaikwad, Shweta Koparde:"AIHumanPoseEstimation:YogaPoseDetection andCorrection",2022
[5] Abhishek Sharma, Yash Shah, Yash Agrawal, Prateek Jain: "Real-Time Recognition of Yoga Poses Using ComputerVisionforSmartHealthCare"inJan2022
[6] Pranjal Sharma, Darshan Pincha, Prateek Jain: "Surya Namaskar: Real-Time Advanced Yoga Pose Recognition AndCorrectionForSmartHealthcare"inSep2022
[7] Josvin Joseand S Shailesh:"Yoga Asana Identification" inMarch2021
[8]SatyamGoyalandAnimeshJain:"Yogaposeperfection using Deep Learning: An Algorithm to Estimate the Error inYogicPoses"inNov2021.
[9] Utkarsh Bahukhandi, Dr. Shikha Gupta: "Yoga Pose Detection and Classification Using Machine Learning Techniques" in International Research Journal of Modernization in Engineering Technology and Science 2021
[10] Ajay Chaudhari, Omkar Dalvi, Onkar Ramade, Dayanand Ambawade: "Yog-Guru: Real-Time Yoga Pose CorrectionSystemUsingDeepLearningMethods"inIEEE, 2021
[11] Deepak Kumar, Anurag Sinha: "Yoga Pose Detection and Classification Using Deep Learning" in International Journal of Scientific Research in Computer Science, EngineeringandInformationTechnology,2020
[12] Shruti Kothari: "Yoga Pose Classification Using Deep Learning"inSanJoseStateUniversityPublication,2020
[13] Ze Wu, Jiwen Zhang, Ken Chen, and Chenglong Fu: "Yoga Posture Recognition and Quantitative Evaluation withWearableSensorsBasedonTwo-StageClassifierand PriorBayesianNetwork";MDPI,Basel,Switzerland,2019
[14] Munkhjargal Gochoo, Tan-Hsu Tan, Shih-Chia Huang, Tsedevdorj Batjargal, Jun-Wei Hsieh, Fady S. Alnajjar, and Yung-Fu Chen: "Novel IoT-Based Privacy-Preserving Yoga Posture Recognition System Using Low-Resolution Infrared Sensorsand Deep Learning";IEEEINTERNET OF THINGSJOURNAL,2019
[15] Muhammad Usama Islam, Hasan Mahmud, Faisal Bin Ashraf, Iqbal Hossain, Md. Kamrul Hasan: "Yoga Posture RecognitionByDetectingHumanJointPointsInRealTime Using Microsoft Kinect"; IEEE Region 10 Humanitarian TechnologyConference(R10-HTC),2017
[16]SharfuddinWaseemMohammed,VigneshGarrapally, Suraj Manchala , Soora Narasimha Reddy and Santosh Kumar Naligenti: "Recognition of Yoga Asana from RealTime Videos using Blaze-pose"; International Journal of ComputingandDigitalSystems,2016
[17] J. Carreira, P. Agrawal, K. Fragkiadaki, and J. Malik.: "Human pose estimation with iterative error feedback."; ComputerVisionFoundation,2016