Gym AI analyzes squat, bicep curl and lateral raise videos with MediaPipe Pose and an LSTM network to flag movement errors that can cause injuries.
An academic project to develop a smart system that analyzes resistance exercises from video and detects movement errors that may lead to musculoskeletal injuries, carried out by a team of students with technical assistance from Techno Enjaz. The application, called Gym AI, extracts a digital skeleton of the body with MediaPipe Pose (the BlazePose model, 33 joint landmarks), then analyzes the movement sequence over time with an LSTM network to judge whether the form is correct or incorrect.
The system covers three common exercises: the squat, the bicep curl, and lateral raises, and works on a live camera feed or a saved video file. Tests showed it can distinguish correct from incorrect form in these exercises under suitable recording conditions with near-real-time response. Its aim is an objective, low-cost tool that helps athletes when no qualified coach is present.
| Item | Details |
|---|---|
| Project type | Desktop application for movement-quality analysis using computer vision and deep learning |
| Field | Computer vision, human pose estimation, temporal modeling, sports safety |
| Project status | Working prototype named Gym AI, tested on examples from the camera and video files |
| Techno Enjaz role | Technical support and assistance to the students during project development |
| Core technologies | Python, MediaPipe Pose (BlazePose), LSTM / LSTM Autoencoder, TensorFlow/Keras, OpenCV, NumPy, Tkinter |
| Outputs | Form classification (correct/incorrect) with a confidence value, skeleton overlay on the image, camera and video-file support |
Many athletes train with resistance exercises without direct supervision from a coach, so they miss errors such as excessive forward knee travel in the squat or loss of balance during a lift. These deviations can accumulate and cause chronic injuries.
The technical difficulty is that an exercise cannot be judged from a single frame: the knee angle may look correct at one moment, while its evolution across the full movement, from the descent to the lowest point and back up, reveals a balance or stability problem. The project therefore built its approach on analyzing the temporal sequence of movement rather than isolated frames, with a focus on prevention before injury occurs.
The project was carried out academically by the students, with technical assistance and support from Techno Enjaz during development. It is presented among the office's work as an example of its support for projects that combine human pose estimation with temporal neural networks in sports and health applications.
Gym AI follows the sequence documented in the project:
The project presents two temporal approaches:
| Aspect | Standard LSTM | LSTM Autoencoder |
|---|---|---|
| Learning type | Usually supervised | Unsupervised |
| Main goal | Classify or predict the movement | Reconstruct the time sequence |
| Need for incorrect examples | Yes | No; training on correct movements is enough |
| Error detection mechanism | Direct classification decision | Reconstruction error rising above a set threshold |
The idea behind the LSTM Autoencoder is that correct movement has a regular temporal pattern the model can compress and reconstruct accurately, while incorrect movement disrupts this pattern and raises the reconstruction error, so it is flagged as an anomaly. This approach helps because collecting enough examples of every kind of error is hard: movement errors are many and varied.
Training data was gathered from three sources:
| Source | Content |
|---|---|
| Videos recorded specifically for the project | People performing the exercises correctly and incorrectly |
| AI-generated videos from PixVerse AI | Specific movement patterns and form errors with controllable type and severity |
| Open datasets from Kaggle | Variety in camera angles, lighting, backgrounds, and body types |
The videos were processed automatically: each video is read frame by frame, landmarks are extracted with MediaPipe, and the time series are saved to a CSV file used to train the model. The project does not state the number of videos or how they were split between training and testing.
| Tool | Role |
|---|---|
| Tkinter and filedialog | GUI and video-file selection |
| OpenCV | Frame reading, color conversion, and pre-processing |
| Pillow | Displaying images in the interface |
| MediaPipe | Extracting body landmarks |
| NumPy | Converting landmarks into numerical arrays |
| TensorFlow and Keras | Building, loading, and running the model |
| Joblib | Saving and loading models |
| threading | Keeping the interface from freezing during processing |
The Gym AI interface consists of a welcome screen with a "Start" button that loads the model and initializes MediaPipe, a screen for choosing the input source (Camera or Video File), and a window that shows the frames and the analysis result, with the option to go back and choose again.
The project presented practical examples from the application:
These are illustrative examples: the project provides no overall numerical metrics such as accuracy on a test set, a confusion matrix, or per-frame processing time. It describes the response as near real time, while noting that it depends on the processing power of the device used.
According to the project's proposals:
The system analyzes video of a person exercising and provides only an assistive assessment of movement quality; it does not replace a qualified coach or medical advice when there is pain or injury.
If you are working on an AI-based sports image and video analysis system and need technical help developing the idea or building a prototype, you can contact the Techno Enjaz office to discuss your project's requirements and the appropriate scope of support.
An academic application built by a team of students with technical assistance from Techno Enjaz that analyzes resistance-exercise video to detect movement errors that may cause musculoskeletal injuries. It works on a live camera or video files and covers the squat, bicep curl, and lateral raises.
MediaPipe Pose, based on BlazePose, extracts 33 joint landmarks from each frame, which become a 99-value vector, turning the video into a time series. An LSTM model then analyzes the whole sequence rather than a single frame, and in the LSTM Autoencoder approach a high reconstruction error marks the movement as anomalous or incorrect.
The data came from videos recorded specifically of people performing the exercises correctly and incorrectly, AI-generated videos from PixVerse AI, and open datasets from Kaggle. All were converted into joint landmarks stored in a CSV file; the project does not state the number of videos or how they were split.
Tests showed correct and incorrect form being distinguished under suitable recording conditions with near-real-time response, such as "Bicep Correct" and "Squat Wrong" outputs with a confidence of 1.00, and successful detection in a home setting and from a frontal angle. The project does not report an overall accuracy on a test set.
It is limited to 2D analysis of three exercises with only a correct/incorrect judgment, and it does not provide medical diagnosis. Joint tracking degrades in poor lighting or when a limb is occluded, depth estimation depends on camera angle, and it may slow down on low-resource devices.
Reach out to Techno Enjaz engineering team to discuss technical feasibility, prototype hardware, and development scope.
Discover more engineering prototypes and academic implementations
A virtual board for writing, drawing, picking colors and erasing with hand gestures in front of a camera, built in Python with MediaPipe hand tracking and OpenCV.
A prototype that detects people with YOLOv8 and tracks them with DeepSORT in defined work areas, counts presence time in seconds, and exports records to Excel.
An exam-monitoring prototype using YOLO, MediaPipe and OpenCV to analyze a student's head direction, flagging head turns in red for review and forward gaze in green.
