HobbyHackers

Inspiration

We were inspired by how often people perform gym exercises incorrectly, leading to injuries or poor results. Hiring personal trainers isn’t always affordable or accessible, so we wanted to create an AI-powered virtual trainer that helps users maintain perfect form anywhere, anytime.

What it does

Our app uses MediaPipe’s pose detection to track a user’s body in real time, analyzing joint angles and posture accuracy. It provides instant visual and audio feedback to correct form during workouts, ensuring every rep is safe and effective.

How we built it

We used Python with MediaPipe for pose tracking, OpenCV for video processing, and NumPy to calculate joint angles using avector geometry. The front end displays live posture tracking, while backend logic compares detected poses to reference datasets of correct exercises. We integrated a simple feedback loop that highlights incorrect postures and guides users to adjust.

Challenges we ran into

The main hurdles were achieving accurate pose estimation under different lighting and camera angles, reducing lag for real-time feedback, and adapting posture thresholds for users of different body types. Fine-tuning MediaPipe’s parameters and optimizing frame processing took significant effort.

Accomplishments that we're proud of

We successfully built a system that delivers reliable posture detection and actionable feedback with minimal delay. Seeing our model recognize and correct squats, lunges, and push-ups accurately was a major milestone.

What we learned

We deepened our understanding of computer vision, body kinematics, and real-time data processing. We also learned how small optimizations in algorithm design can drastically improve performance and user experience.

What's next for HobbyHackers

We plan to expand the project into a full fitness assistant with personalized workout plans, progress tracking, and integration with wearable devices for heart rate and calorie monitoring. Our goal is to make smart fitness guidance accessible to everyone.

Built With

• bertmodel
• computervision
• mediapipe
• python