Easy Coach

Status: Published · Year: 2022 – 2023 · Stack: OpenPose · Python · Real-time, webcam-only

From 2020 to 2022, the COVID-19 pandemic made normal outdoor activities difficult. To help people maintain physical health at home, I built Easy Coach — a posture coach that needs nothing more than a laptop webcam. A full interactive write-up (with the original table of contents) is available at the original project page.

Main coaching algorithm

Home workouts (squats, push-ups, lunges, shoulder press, overhead tricep extensions) require monitoring joint angles and identifying repeated movement phases. Easy Coach uses extrema in joint-angle-vs-time graphs — built from OpenPose keypoints — to find phases, count repetitions, and estimate movement speed.

OpenPose outputs 2D keypoint coordinates. Choosing three joints forms a triangle; the law of cosines then converts side lengths into a joint angle in [0, π], which is robust for downstream processing.

Recording every per-frame angle caused memory and I/O overhead, so Easy Coach instead detects local maxima/minima online, in-stream, and aggregates only those representative values — constant time per frame, no memory or I/O bottleneck. Local-extrema detection relies on sign changes in the derivative of the angle-time signal, which is lightweight and real-time-friendly.

Repetitive exercises produce repeating extremum patterns (e.g., knee angle during squats). Tracking derivative sign patterns and aligning extrema determines the current movement phase, repetition index, and when to stop recording.

Accuracy: handling OpenPose misdetections

During beta testing, most coaching inaccuracies traced back to OpenPose misdetections caused by occlusion — another person or moving object behind the user, or furniture partially blocking a body part.

OpenPose marks fully missing values with -1, which are simply discarded. The harder case is an obstacle mistaken for a body part, producing a plausible angle (still within 0°–180°). Easy Coach flags and corrects these with two criteria:

• Criterion 1 — an angle below the exercise-specific physical lower bound is marked erroneous.
• Criterion 2 — a frame-to-frame angular change exceeding a plausible physical threshold (derivative-based) is marked erroneous.

Flagged frames are replaced with the average of previous valid observations, smoothing the signal and improving coaching reliability.

Feedback from specialists

Presentation feedback from industry and academia emphasized two points: (1) UI/UX matters — gamification and engagement features are needed to retain users; (2) prior-art research would sharpen differentiation and product strategy. Both shaped the subsequent roadmap.