Search for a command to run...
Body posture recognition plays a vital role in applications such as healthcare monitoring, fitness training, and human-computer interaction. Despite advances in lightweight pose estimation frameworks, existing systems struggle with computational overhead, occlusion handling, and limited accuracy under variable conditions. This research introduces an intelligent posture correction system that integrates Google's MediaPipe framework with three advanced machine learning architectures: XGBoost-LightGBM ensemble, 1D Convolutional Neural Networks, and lightweight Transformer models. The system performs high-fidelity pose estimation by detecting 33 key body landmarks through RGB camera input using MediaPipe's BlazePose model, enabling detailed 2D and 3D skeleton mapping optimized for real-time operation on mobile and desktop platforms without cloud dependency. The primary contribution lies in the hybrid algorithmic approach combining gradient boosting ensembles for feature-based classification, temporal convolutional networks for sequential pattern recognition, and attention mechanisms for landmark relationship modeling. Our system accurately identifies critical postures including standing, sitting, bending, and stretching through geometric joint angle calculations and relative distance metrics. The XGBoostLightGBM ensemble achieves 97.8% accuracy, the 1D CNN model reaches 98.2%, while the Transformer-based architecture attains 98.7% classification accuracy, significantly outperforming conventional k-NN (92.4%) and SVM (95.6%) approaches. The system incorporates adaptive filtering mechanisms to handle partial occlusion, dynamic lighting variations, and non-standard body orientations effectively.