Providing automated pose correction feedback for at-home fitness solutions remains a big challenge. This work introduces PosePilot, a novel integrated solution evaluated on Yoga, an ancient Indian practice rooted in holistic well-being. PosePilot has been trained on real-world videos sourced from Yoga practitioners, encompassing the performance of six asanas captured from four different viewpoints. Vanilla LSTM was employed for sequential modeling, enabling PosePilot to capture temporal dependencies for pose recognition. Furthermore, we utilized BiLSTM with Multihead Attention to enhance the model’s capacity to examine both forward and backward temporal contexts, enabling selective focus on pertinent limb angles for error detection. Finally, we also provide correctives for each pose for every temporal stage of the pose.
Our inhouse was created with videos from four angles, featuring 14 participants (ages 17-25) performing six poses. Frame-by-frame keypoint extraction using Mediapipe identified 33 keypoints, with 17 used to compute 680 angles for pose analysis. The dataset contains 336 videos, filmed indoors with controlled lighting.
