AI Vision Models for Smart Padel Courts

Job ID: 40554428

Budget: $2 – $8 USD

We turn padel courts into smart courts. Two cameras watch each match; players open a shared page afterward to see their highlight clips and performance stats. We're hiring a freelancer to build the computer-vision models that make this work.

The core challenge is the vision pipeline. From two camera feeds, you'll detect and track four players, hold a stable identity for each across a full match, map their movement into real court coordinates, and turn that motion into meaningful output — highlight clips and per-player stats. You own the models and the data they produce; our team builds the player-facing page on top.

What you'll build:


Player detection + multi-object tracking — YOLO-class detection with ByteTrack/DeepSORT-style tracking, maintaining stable per-player identity across occlusion and crossover
Pose estimation — for gesture-based highlight triggers (e.g. a player holding a T-pose)
Court mapping — homography from image plane to real court coordinates (OpenCV); auto-calibration from court lines is a plus
Movement analytics — from ~5 Hz position tracks: heatmaps, zone-time, distance, and derived spatial stats
Highlight extraction — 30-second clips triggered per side, packaged into a clean, documented per-match output (JSON + clips + heatmaps)


Exploratory (bonus): rally segmentation and shot detection from motion — approaches open, best-effort.

You should be strong in: modern object detection & multi-object tracking, pose estimation, camera calibration / homography, and turning noisy real-world tracking data into robust metrics. Python + OpenCV throughout.

Deployment context: the pipeline runs on an NVIDIA Jetson processing real-time video, so comfort optimizing models for constrained hardware and working with RTSP/GStreamer is valued — the ML is the heart of the role, but you'll build it on the device. You should be at home working on a headless Linux box remotely over SSH (no GUI), moving code and files via scp, and running/debugging everything from the command line. Familiarity with the Jetson/JetPack stack (CUDA, TensorRT) is a plus.

The rig is already built and remotely accessible; a full technical handoff goes to shortlisted candidates.

To apply: share relevant CV/ML work and a short note on how you'd approach the detection → tracking → court-mapping chain