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

TL;DR

PosePilot addresses the need for real-time, personalized posture correction in AI-driven fitness by integrating pose recognition with per-frame corrective feedback. It introduces an edge-friendly architecture that uses a Vanilla LSTM for recognition and a BiLSTM with multi-head Attention for sequence-based corrections, trained on an in-house Yoga dataset with 680 joint-angle features reduced to 9 correction angles. The system achieves a recognition accuracy of $97.52\%$ and a correction forecasting MSE of $0.00138$, while running on a Raspberry Pi 4 with quantization that incurs only about $1\%$ accuracy loss. This work demonstrates the practicality of end-to-end edge deployment for real-time posture guidance and provides a foundation for extending to physiotherapy and sports coaching with improved explainability.

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.

Figures (4)

• Figure 1: Sample of In-house Dataset.
• Figure 2: 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. A BiLSTM then analyzes those frames to forecast the next joint-angle vector $\hat{p}_{t}$. These deviations are processed by an error-calculation module, and the user interface presents per-angle correction feedback alongside a dynamic visualization of pose improvement.
• Figure 3: 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. A correctly performed Utkatasana is shown alongside for comparison.
• Figure 4: Pose Recognition and Personalized Corrective Feedback GUI