Real-Time Computer Vision & Distance Estimation

From pixels to metres

A detector that finds a basketball rim in a frame is useful. A system that also reports how far away that rim is, from one ordinary camera, is the difference between an object detector and a piece of perception a robot can act on.

This project trains the detector, derives the geometry, and puts both into a real-time video pipeline that draws its own answers back onto the frame.

Trained from scratch, not fine-tuned

• Trained a custom YOLOv5 model from scratch on a labelled dataset built for this task, rather than adapting an off-the-shelf checkpoint.
• Runs in real time on GPU, detecting basketball rims in a live feed.

The geometry that makes one camera enough

A monocular camera cannot measure depth directly — but a target of known physical size gives it back. The apparent size of the rim in pixels falls off predictably with distance, so a calibrated pixel-to-real-world distance formula converts box geometry into a metric estimate.

Calibration is what makes it trustworthy: the relationship is fitted against known distances rather than assumed from lens specifications.

Everything drawn back onto the frame

• A real-time video pipeline overlaying bounding boxes, confidence scores and the computed distance on each frame.
• Annotated video export, so a run can be reviewed, shared and checked against ground truth after the fact.
• Built on OpenCV for capture, drawing and encoding, with CUDA keeping inference ahead of the frame rate.

Running on a real court

Unedited output. The detector holds the rim through camera movement, occlusion from branches and a change of backdrop, and the distance estimate updates on every frame — 4.75 m in the opening shot.

Source

Code on GitHub.