Yoga Practice Posture and Performance Feedback from Machine Vision ^†

Abstract

Recently, the yoga industry has witnessed rapid technology integration, particularly through computer vision and machine learning methods, to improve the practice experience. We developed a yoga posture detection and feedback system using a Raspberry Pi 5-based hardware setup and a combination of the Open Source Computer Vision Library, MediaPipe, and TensorFlow. The system records videos of practitioners doing five standard yoga poses and provides rapid data-based feedback to help fix misalignments. The model accurately classifies poses using convolutional neural networks and identifies important joint landmarks and limb angles, which is confirmed through confusion matrix analysis. The developed system lets its users practice safely without direct supervision from an instructor. It increases access to practical yoga training. The setup also allows for future physical therapy and athletic training applications, where proper form and alignment matter.

1. Introduction

Yoga has gained popularity due to its many physical and mental health benefits. An essential part of yoga practice is proper body alignment, which ensures effectiveness and safety and helps prevent injuries. Traditionally, instructors guide yoga sessions and provide real-time corrections and feedback on the practitioners’ posture. However, not everyone can attend in-person classes or hire private yoga instructors, especially in today’s fast-paced digital world [1]. To meet this need, technological solutions are used to help practitioners maintain correct form and posture during yoga. Advances in computer vision have also led to systems that automatically detect and classify yoga poses. These systems assess posture by analyzing joint positions, limb angles, and body symmetry, providing immediate feedback without requiring human supervision [2]. MediaPipe v0.10.33, TensorFlow 2.21.0, and Open Source Computer Vision Library (OpenCV) are widely used tools for pose estimation and classification. When implemented on affordable hardware like the Raspberry Pi 5, these technologies enable and allow for accessible and low-latency yoga training [3].

Recent research has investigated the use of computer vision and deep learning methods aimed at recognizing and correcting yoga postures. Multiple studies [4,5,6,7] emphasize real-time detection of yoga poses, reconstruction of postures, and feedback mechanisms, showcasing the efficiency of skeletal tracking and feature-based models in assessing body alignment. These methods facilitate precise pose identification and corrective input, demonstrating encouraging outcomes in enhancing user performance and assisting autonomous yoga practice without direct guidance from an instructor.

Furthermore, additional research [8,9,10,11] has explored posture monitoring and classification through machine learning utilizing different methods, such as human pose synthesis, IoT-driven posture monitoring systems, and convolutional neural network models for recognizing activities and postures. These methods have also been applied to various fields [12,13,14,15,16,17], including medical image categorization, supportive technologies, agricultural assessment, and object recognition. The ongoing effectiveness of convolutional neural networks and computer vision techniques in these areas emphasizes their resilience and flexibility in managing intricate visual data, endorsing their use in yoga posture evaluation systems.

In this study, to help users achieve correct yoga postures, we developed a system that assesses a practitioner’s posture by applying and training the system using TensorFlow, MediaPipe, and OpenCV. The developed system was tested and evaluated for performance using a confusion matrix.

2. Methodology

Figure 1 shows the workflow of the developed yoga posture detection and feedback system. A Raspberry Pi with a camera module captures live images of practitioners performing yoga poses. Using OpenCV and MediaPipe, skeletal landmarks and joint angles are extracted and passed to a convolutional neural network (CNN) in TensorFlow for classification. Feedback is provided visually on a monitor and audibly through a speaker.

System Component

The system’s component is the Raspberry Pi 5 Model B. An 8 megapixel Raspberry Pi camera module is adopted to capture images. A connected monitor displays posture feedback through a high-definition multimedia interface, while a speaker provides auditory feedback (Figure 2). The system is compact, portable, and efficient for real-time processing.

OpenCV is used for image capture and preprocessing, MediaPipe extracts skeletal landmarks, and TensorFlow’s CNN performs classification. Python 3.13.13 then integrates these components for smooth data flow and feedback delivery (Figure 3).

The system was tested in a controlled environment. The camera was placed two meters away at a slightly elevated angle to capture the full body. Figure 4 shows the experimental setup.

Figure 5 presents the five yoga poses included in the dataset: downward-facing dog, tree, warrior, plank, and goddess.

Table 1. Dataset of yoga poses used in the system.

Pose	Number of Samples
Downward-facing dog	30
Tree	30
Warrior	30
Goddess	30
Plank	30

summarizes the data collected for five yoga poses, with 30 samples per pose used in training and testing.

3. Results and Discussion

The system achieved robust classification results. Figure 6 shows the sample output of the developed system identifying the warrior pose.

A confusion matrix was generated to evaluate accuracy (

Table 2. Confusion matrix for yoga pose classification.

Actual Pose	Predicted Pose
	Downward-Facing Dog	Tree	Warrior	Goddess	Plank	Total
Downward-facing dog	28	0	0	1	1	30
Tree	0	29	1	0	0	30
Warrior	1	0	27	1	1	30
Goddess	0	1	1	28	0	30
Plank	2	0	0	0	28	30
Total	31	30	29	30	30	150

and

Table 3. Predicted yoga posture.

Total Yoga Posture Prediction	Predicted Yoga Posture
150	Correctly Predicted Yoga Posture	140
	Incorrectly Predicted Yoga Posture	10

).

Out of 150 total samples, 140 were correctly predicted, achieving 93.33% accuracy. These results indicate that the CNN model, combined with MediaPipe landmark extraction, was able to effectively distinguish between the selected yoga postures under the given test conditions.

The accuracy of the model was computed using the confusion matrix results using Equation (1). The total correct classifications were 140 out of 150, resulting in an overall model accuracy of 93.33%.

$$Accuracy={\displaystyle \frac{{\displaystyle {\sum }_{n=1}^5}{A}_{nn}}{{\displaystyle {\sum }_{i=1}^5}{\displaystyle {\sum }_{j=1}^5}{A}_{ij}}}$$

(1)

An analysis of misclassifications reveals that errors were minimal and primarily occurred between poses with similar structural characteristics. For example, warrior was occasionally misclassified as goddess or plank, likely due to similarities in lower-body stance and limb positioning. Similarly, plank and downward-facing dog showed minor confusion, which may be attributed to transitional frames or slight variations in hip alignment during data capture. Despite these minor overlaps, the model maintained high per-class consistency, with no class dropping below 90% accuracy. These findings demonstrate that landmark-based feature extraction combined with CNN classification is effective for real-time embedded pose recognition. The results support the feasibility of deploying such systems in practical applications such as home-based yoga guidance, rehabilitation monitoring, and fitness training, where reliable but lightweight models are required.

4. Conclusions and Recommendations

This study set out to create a real-time yoga posture feedback system using machine vision on an affordable embedded device. The system ran on a Raspberry Pi 5 and used MediaPipe for detecting body landmarks, OpenCV for image processing, and a Convolutional Neural Network (CNN) to classify poses. The model used landmark coordinates as input, and rule-based evaluation provided real-time corrective and motivational feedback. Tests showed the system could accurately recognize five common yoga poses, reaching an overall accuracy of 93.33%. This result shows that lightweight deep learning models can work well for real-time pose recognition on edge devices.

These results show that embedded systems could make yoga training, physical therapy, rehabilitation, and sports monitoring more accessible, especially at home or remotely. In the future, the pose dataset should include more complex and dynamic movements. The system should also be improved to work better with different body types and environments, use higher-resolution or depth sensors, support smoother pose transitions, and offer a mobile app to help users track their progress over time.