Posture Identification and Evaluation of Weight-Lifting Workout Styles using LSTM Networks

Abstract

Proper posture during weightlifting exercises is a critical factor for maximizing muscular gains, enhancing metabolic performance, and minimizing the risk of injuries to vulnerable areas such as the lower back, shoulders, elbows, and knees. However, many fitness enthusiasts, particularly those training at home without professional supervision, lack access to reliable feedback mechanisms that can ensure safe and ef fective exercise practices. Incorrect posture, even when subtle, not only reduces training efficiency but also significantly increases the likelihood of long-term musculoskeletal problems. Previous research efforts to ad dress this challenge have largely relied on sensor-based systems such as Inertial Measurement Units (IMU) and Electromyography (EMG). While these methods provide high levels of accuracy, they remain costly, intrusive, and highly dependent on correct sensor placement, which limits their scalability and practicality for home-based users. On the other hand, vision-based solutions such as Kinect and MediaPipe have shown promise but are often restricted to static frame analysis. These approaches fail to capture the temporal dynamics of motion, making them less effective for analyzing continuous weightlifting sequences or deliver ing detailed, joint-specific corrections. This study addresses these limitations by developing a non-intrusive, video-based posture evaluation system specifically for weightlifting exercises, focusing on Barbell Biceps Curls and Tricep Kickbacks. The system leverages Google’s MediaPipe to extract skeletal keypoints corresponding to major joints such as shoulders, elbows, wrists, hips, and knees. From these landmarks, critical movement features including joint angles, torso inclination, and head orientation are derived and normalized. These temporal features are then processed using a Long Short-Term Memory (LSTM) network, which is capable of modeling sequential motion patterns across frames. This enables the classification of correct versus incorrect posture with high accuracy, achieving 94% for Biceps Curls and 96% for Tricep Kickbacks. To make the system practical and accessible, a lightweight web application was implemented, allowing users to upload workout videos and receive frame-by-frame analytical feedback. The application highlights deviations such as flared elbows, excessive torso rotation, or improper wrist alignment, offering users actionable guidance to improve form. Unlike expensive sensor-based systems, this approach requires only a standard smartphone camera, making it highly suitable for home fitness users. By integrating vision-based pose estimation with temporal deep learning, this study introduces a novel framework that not only improves workout safety and effectiveness but also sets a new benchmark for intelligent, accessible, and low-cost digital coaching in fitness technology.