PosePilot: An Edge-AI Solution for Posture Correction in Physical Exercises

Abstract: Automated pose correction remains a significant challenge in AI-driven fitness systems, despite extensive research in activity recognition. This work presents PosePilot, a novel system that integrates pose recognition with real-time personalized corrective feedback, overcoming the limitations of traditional fitness solutions. Using Yoga, a discipline requiring precise spatio-temporal alignment as a case study, we demonstrate PosePilot's ability to analyze complex physical movements. Designed for deployment on edge devices, PosePilot can be extended to various at-home and outdoor exercises. We employ a Vanilla LSTM, allowing the system to capture temporal dependencies for pose recognition. Additionally, a BiLSTM with multi-head Attention enhances the model's ability to process motion contexts, selectively focusing on key limb angles for accurate error detection while maintaining computational efficiency. As part of this work, we introduce a high-quality video dataset used for evaluating our models. Most importantly, PosePilot provides instant corrective feedback at every stage of a movement, ensuring precise posture adjustments throughout the exercise routine. The proposed approach 1) performs automatic human posture recognition, 2) provides personalized posture correction feedback at each instant which is crucial in Yoga, and 3) offers a lightweight and robust posture correction model feasible for deploying on edge devices in real-world environments.

• The paper introduces a two-stage system that first recognizes Yoga poses using a Vanilla LSTM and then corrects errors with a BiLSTM augmented by multi-head attention.
• The methodology achieves 97.52% recognition accuracy and a mean square error of 0.00138, highlighting its precision in forecasting pose trajectories.
• The lightweight design enables real-time corrective feedback on edge devices, promising wider applications in physiotherapy and sports coaching.

PosePilot: An Edge-AI Solution for Posture Correction in Physical Exercises

Introduction

"PosePilot: An Edge-AI Solution for Posture Correction in Physical Exercises" (2505.19186) introduces a novel methodology combining pose recognition with real-time personalized corrective feedback, primarily focusing on Yoga postures. This approach aims to enhance the precision of physical exercise execution while being optimized for deployment on edge devices. The system integrates a Vanilla LSTM for capturing temporal dependencies in pose recognition and a BiLSTM with multi-head Attention for detailed error detection, maintaining computational efficiency and delivering instant corrective feedback.

Methodology Overview

The core of PosePilot is its computational architecture, designed to handle the spatio-temporal complexities of Yoga movements. It utilizes Long Short-Term Memory (LSTM) networks to leverage sequential data, which is crucial for capturing the nuances inherent to precision-focused exercises such as Yoga. By focusing on key joint angles and employing a lightweight computational model, PosePilot is well-suited for real-time application on resource-limited devices, thus offering practical solutions for individuals practicing at home or outdoors.

Figure 1: PosePilot Overview. Video frame's extracted joint angles from time [0,t] are first fed into an LSTM that captures motion and classifies the current yoga pose, selecting one of six trained correction models.

Pose Recognition and Correction

PosePilot's methodology involves a two-stage process: pose recognition followed by corrective feedback. The system first identifies the Yoga asana through a Vanilla LSTM that processes videos of poses, isolating key frames that reflect significant posture changes using angle deviation calculations. Upon recognizing the pose, a BiLSTM model analyzes temporal data across joints to predict the following joint-angle vector, flagging deviations that exceed predefined thresholds.

Figure 2: Correction graph for an incorrectly performed Utkatasana. Error corresponding frames are marked with red crosses. Any joint angle deviating more than 1.5 standard deviations from the ideal pose is flagged, and red vectors show the adjustment needed to bring each point back within the acceptable range.

Dataset and Evaluation

PosePilot was tested on an in-house dataset specifically curated for Yoga, featuring high-quality video sequences capturing poses from multiple angles. This dataset overcomes existing limitations such as poor lighting and participant diversity. The model achieved an average recognition accuracy of 97.52% across multiple asanas, and a mean square error (MSE) of 0.00138 for corrections, demonstrating exceptional ability to forecast pose trajectories and suggest accurate modifications.

Figure 3: Sample of In-house Dataset.

Practical Implications and Future Directions

PosePilot presents significant implications for improving exercise routines by reducing injury risk through precise real-time guidance. Its real-world applicability is enhanced by edge deployment, which ensures privacy and independence from network constraints. The system's lightweight architecture further advocates for broader use in settings like physiotherapy and sports coaching.

Figure 4: Pose Recognition and Personalized Corrective Feedback GUI.

Conclusion

The "PosePilot" paper outlines a sophisticated system that significantly advances pose recognition and correction capabilities within fitness applications. While currently optimized for Yoga, its foundational model can be readily adapted to other exercise forms. Future work will explore optimization of feature extraction latency and deeper real-time deployment across various physical activity domains.

PosePilot establishes a new benchmark for integrating AI into exercise routines, marking a step forward in personalized fitness training solutions on edge devices. The potential for extension into diverse fields, such as rehabilitation, indicates the promising trajectory for this line of research.