How Can AI Teach You to Lift Safely Without Getting Hurt?

The deadlift is one of the most loved exercises in the gym, but a wrong posture can turn a workout into a source of injuries. A research team therefore created an app that, using deep learning, analyzes lifters' videos and recognizes three deadlift styles – conventional, Romanian and sumo – while also flagging common mistakes like a rounded back, overextension, or premature hip lift. To train the algorithm, they recorded hundreds of side‑view videos and carefully annotated them. They then adapted the MoveNet model to track key joint points, turning them into coordinates that describe the movement. These coordinates fed both a convolutional neural network (CNN) and an LSTM model, which excels at recognizing temporal sequences. The results are striking: both models achieved an F1 score of 0.99 for style classification and even 1.00 for correct form detection. The app not only shows the right poses but also lets users replay their video with overlaid keypoints, providing instant feedback. In practice, it works like a digital personal trainer that warns you before your back rounds, helping you improve technique and lower injury risk.

The MoveNet model, originally designed for real‑time human pose tracking, was fine‑tuned to extract 17 anatomical keypoints (ankle, knee, hip, shoulder, etc.). These points were normalized relative to torso length to make the model independent of the athlete's stature. The CNN employed a modified ResNet‑18 architecture, with 3×3 convolutional layers and batch normalization, to capture spatial features of the keypoints in individual frames. In parallel, the LSTM model, with two layers of 128 units each, processes sequences of 30 frames (about one second of video) to learn the temporal dynamics of the movement. Combining both architectures enables detection of not only static posture but also the speed and synchronization of body segments. For error classification, a threshold system based on angular deviations was implemented: for instance, a lumbar angle below 70° for more than 0.2 s is flagged as "rounded back". Training ran on an NVIDIA RTX 3080 GPU, with data augmentation (rotations ±10°, zoom 0.9‑1.1) to enhance robustness. The result is a model capable of real‑time inference (~30 fps) on mobile devices, paving the way for AI‑driven sports coaching solutions.