AI-Powered Real-Time Fitness Tracking and Posture Detection System

Abstract

Maintaining correct exercise form is essential for fitness effectiveness and injury prevention, yet most existing fitness applications focus solely on quantitative metrics such as step count, calories burned, and heart rate, ignoring qualitative aspects of movement correctness. This work introduces an AI-driven real-time fitness monitoring and posture analysis system that utilizes computer vision and deep learning techniques to evaluate exercise performance through a regular webcam without requiring wearable devices. The proposed framework uses MediaPipe Pose Estimation to identify 33 key body points in every video frame for real-time skeletal tracking. These pose features are processed by a CNN–LSTM hybrid model, where the Convolutional Neural Network (CNN) extracts spatialposture characteristics, while the Long Short-Term Memory (LSTM) module learns sequentialmovement patterns over time. The integrated pipeline automatically classifies exercises such as squats, push-ups, and lunges, counts repetitions accurately, and computes joint angles to identify postural deviations.Upon detecting incorrect form, the system delivers instant corrective feedback via visual alerts or voice prompts, mimicking the role of a personal trainer. Operating at over 30 FPS, the system ensures smooth real-time monitoring.Experimental evaluation demonstrates high exercise classification accuracy and effective posture correction, making professional-quality fitness coaching accessible and affordable to all users regardless of location or resources.

Introduction

This paper presents an AI-Based Real-Time Fitness Tracker and Posture Detection System that uses computer vision and deep learning to monitor exercise performance without requiring wearable devices. Traditional fitness applications mainly track quantitative metrics such as calories burned, distance, and workout duration, but they cannot assess whether exercises are performed correctly. Incorrect posture during workouts often leads to injuries, muscle strain, and long-term physical issues.

The proposed system uses a standard webcam and MediaPipe BlazePose to detect 33 body landmarks in real time. A CNN–LSTM hybrid model analyzes body movements to recognize exercises such as squats, push-ups, and lunges, while also counting repetitions. Joint-angle calculations are used to identify posture errors, and the system provides immediate corrective feedback through visual alerts or voice prompts. The entire process operates at over 30 frames per second, offering users a virtual coaching experience similar to professional fitness guidance.

The literature review highlights advances in human activity recognition, pose estimation, and deep learning techniques such as OpenPose, BlazePose, CNNs, and LSTMs. Existing solutions often rely on wearable sensors, lack real-time feedback, or address only one aspect of fitness monitoring. The proposed system integrates exercise recognition, repetition counting, posture analysis, and corrective feedback into a single framework.

The methodology follows a ten-step pipeline: video capture, frame extraction, pose estimation, preprocessing, CNN feature extraction, LSTM-based temporal analysis, exercise classification, repetition counting, posture evaluation, feedback generation, and output display. The software stack includes MediaPipe, TensorFlow/Keras, OpenCV, Flask, React, MongoDB, and Python, while testing was performed on standard desktop and Android hardware.

Experimental results demonstrate strong performance, achieving 94.3% overall exercise classification accuracy, a repetition counting error of only 0.4 repetitions per set, and 91% precision in posture correction detection. The system maintained an average speed of 32 FPS, ensuring smooth real-time operation. The study concludes that combining CNN and LSTM models with pose estimation provides an effective, privacy-preserving, and accessible fitness monitoring solution. Future work includes expanding support for additional exercises and multi-camera setups.

Conclusion

This paper presented an AI-based real-time fitness trackerand posture detection system that combines MediaPipe Pose Estimation with a CNN–LSTM deep learning model to deliver professional-quality exercise monitoring using only a standard webcam. The system automatically recognizes exercises, counts repetitions, evaluates joint-angle-based posture correctness, and providesinstantcorrectivefeedbackatover30FPS.Experimental resultsdemonstrateaclassificationaccuracyof94.3%,repetition counting error of 0.4 reps, and posture correction precision of 91%.By requiring no wearable hardware and running entirely on-device, the system makes personalized fitness coaching accessible and affordable to a wide range of users, bridging the gap between home workouts and professional training. The system also demonstrates strong real-time responsiveness and scalability, which makes the framework practical for both web- based and mobile fitness platforms.

References

[1] Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes Ortiz, “A Deep Learning Approach for Human Activity Recognition using Smartphones,” in Proceedings of the European Symposium on Artificial Neural Networks (ESANN), 2013. [Online]. [2] Available: https://www.esann.org/sites/default/files/proceedings/legacy/es2013-84.pdf [3] N. Y. Hammerla, S. Halloran, and T. Plötz, “Deep, Convolutional, and Recurrent Models for Human Activity Recognition using Wearables,” in Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2016. [Online]. [4] Available: https://www.ijcai.org/Proceedings/16/Papers/220.pdf [5] Y. Zheng, J. Yu, and X. Zhou, “Pose Estimation for Fitness Applications using CNNs,” in IEEE Conference onComputer Vision, 2018. [6] Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh, “Realtime Multi-Person 2D Pose Estimation using Part AffinitFields,”in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. [Online]. [7] Available: https://openaccess.thecvf.com/content_cvpr_2017/papers/Cao_Realtime_Multi-Person_2D_CVPR_2017_paper.pdf [8] F.J. Ordóñez and D. Roggen, “Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition,” Sensors, vol. 16, no. 1, p. 115, 2016. [Online]. [9] Available:https://www.mdpi.com/1424-8220/16/1/115 [10] H.Kaur and M. Singh, “AI-Powered Personal Trainer: A Review,” International Journal of Computer Science and Mobile Computing (IJCSMC), 2021. [11] Y. Chen et al., “Intelligent Personal Trainer: Real-time Exercise Posture Detection and Corrective Feedback,” IEEE Access, 2021. [12] V.Bazarevskyetal.,“BlazePose:On-deviceReal-time BodyPoseTracking,”2020.[Online]. Available:https://arxiv.org/abs/2006.10204

Copyright

Copyright © 2026 Prof. Veda Shree Y C, Bindushree S, Brinda Hemanth, Dhanyashree D, Chandrika G M. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.