Inspiration
Physiotherapy exercises are often done incorrectly at home, leading to slow recovery or even new injuries. We wanted to build something that gives instant, intuitive feedback without expensive hardware — something anyone with a laptop can use. That’s how PsyFlow was born: a real-time physiotherapy coach that overlays a “ghost” demonstration model and guides users toward perfect form.
What it does
PsyFlow lets users follow along with a demo video using a live, overlaid skeleton: Extracts the full-body pose sequence from a physiotherapy demo video Displays a “ghost” skeleton performing the exercise Tracks your body through the webcam Shows your live skeleton beside (and overlapping) the ghost Computes similarity scores for posture and movement Highlights the worst-aligned joint Gives simple coaching cues (“Raise your left arm”, “Straighten your right knee”) Warns when parts of the body go out of frame It effectively turns your laptop webcam into a physiotherapy form coach.
How we built it
Used MediaPipe Pose to detect 33 joints in each frame Built a custom pipeline to extract normalized ghost pose sequences from reference videos Implemented a pose normalization system to compare relative limb positioning independent of camera setup Created a joint-distance similarity metric to score how well a user matches the ghost Added a coaching engine that examines the worst joint deviation and maps it to natural movement hints Built a real-time overlay: yellow ghost skeleton + green live skeleton Optimized everything to run at 30+ FPS in Python on a standard laptop
Challenges we ran into
MediaPipe landmarks can be unstable or missing when limbs leave the frame Full-body similarity metrics were too forgiving at first, then too strict Building stable normalization across different body sizes and camera distances Extracting consistent motion cycles from demo videos Designing feedback that’s meaningful but not overwhelming Keeping real-time performance smooth despite heavy processing Accomplishments that we're proud of Fully working ghost-overlay physiotherapy system Real-time skeleton tracking that’s smooth and visually clean Intelligent feedback that pinpoints the exact joint that needs fixing The system adapts automatically to any exercise video — no retraining required It feels surprisingly intuitive to match your body to the ghost
Achieved this in less than 48 hours at a hackathon
What we learned How difficult pose comparison is beyond simple angle math How to normalize poses across people of different body proportions How to smooth jitter and compensate for incomplete landmark detections That intuitive visual feedback (ghost overlay) is far more effective than just numbers The value of balancing technical accuracy with usability and clarity
What's next for PsyFlow
Add repetition counting, rep timing, and form scoring Allow users to upload their own physiotherapy routines and create custom ghosts Introduce AI-written coaching feedback personalized to the user’s errors Build a web version using TensorFlow.js so it runs entirely in the browser Integrate physiotherapist-approved exercise libraries Add progress tracking & streaks for long-term rehabilitation Build mobile support with AR overlays
Log in or sign up for Devpost to join the conversation.