Computer Vision Exercise Technique Evaluation System

Challenge

Our client, a fitness club, wanted to introduce a new service for its members: a subscription-based solution for evaluating exercise technique. The goal was to support clients who cannot afford a personal trainer.

This service needed to be more affordable than one-on-one training while still increasing overall membership revenue through a recurring subscription model. Additionally, it aimed to improve the quality of independent workouts and reduce the risk of injuries by identifying technique errors in a timely manner.

Our task was to develop a system capable of accurately evaluating exercise execution.

Solution

The user records themselves on video during a workout, and our service automatically analyzes the technique, identifies mistakes, and generates recommendations.

In the uploaded video, the algorithms detect the body position at the key (“peak”) phase, determine the coordinates of major joints (elbows, knees, shoulders, etc.), calculate joint angles, and compare them with reference values for each exercise.

We implemented support for 13 basic exercises. All measurements are performed with the user positioned in a fixed way relative to the camera, ensuring standardization of results.

After processing the video, the user receives:

• An image with overlaid skeletal keypoints
• Screenshots capturing peak positions
• An automated breakdown of mistakes and technique recommendations

Technologies and Architecture

The system is based on the MediaPipe model, which provides highly accurate detection of body keypoints without requiring additional training. For more complex scenarios where MediaPipe’s standard capabilities are not sufficient (for example, precise estimation of distance to the floor), we implemented an additional model that analyzes full-body positioning.

Technical Challenges and Solutions

• To improve recognition stability, we fine-tuned MediaPipe parameters and standardized camera positioning.
• For exercises involving occlusion of body parts, we implemented specialized smoothing and filtering algorithms for unstable keypoints.
• The lack of scale in video was addressed using relative measurements and human anatomical proportions.

Results

The system is already used by the client for individual training sessions and corporate fitness programs. The analysis time for a single exercise takes only a few seconds, and the accuracy of technique evaluation is comparable to that of a professional trainer.

In the near future, we plan to expand the list of supported exercises and integrate the system with personal fitness platforms. The project is ready to scale into a full-featured AI-powered personal fitness platform.