RallyLens
1 camera broadcast-style shot, bounce and speed tracking. No rig, no wearables.
- Client
- Confidential
- Year
- 2025
- Category
- Computer Vision
- Role
- Design + Build
- Timeline
- Ongoing
The problem
Coaches review sessions by eye, so progress is a matter of memory. A busy court makes it worse: a coach and several students at once, a fast small ball, players blocking each other, and wall drills where “on target” is a judgment call. The studio wanted per-player, per-shot data from an ordinary camera on a tripod, not a fixed multi-camera rig and not wearables on every player.
What we built
RallyLens detects and pose-tracks every player, colours them by role, and follows the ball frame to frame as a trajectory spline. A camera-homography step maps the court and the wall targets into real-world coordinates, so bounces and wall hits report distance to target in centimetres and ball speed comes from calibrated displacement. An event timeline marks every bounce, contact and on-target hit; a bird’s-eye minimap shows court position. It runs on recorded clips and exports a per-player shot log.
The result
Every session is now a timeline instead of a memory. Coaches compare students on the same numbers, and wall-drill accuracy is measured in centimetres rather than argued about. The studio is folding it into regular coaching reviews.
- wall-target accuracy, measured rather than argued about
- cmwall-target accuracy, measured rather than argued about
- per-shot ball speed with a session peak
- km/hper-shot ball speed with a session peak
- shot, bounce and rally counts, exportable
- per playershot, bounce and rally counts, exportable
Built with
- Python
- PyTorch
- YOLO
- TrackNet
- OpenCV
- Homography
We deliver what we commit.
Tell us what you're trying to build.
We'll come back within 24 hours with honest feedback on scope, timeline and cost, whether or not we turn out to be the right fit.