Object Detection for Inventory Stocking

The Problem

Counting stock by hand is tedious and easy to get wrong, and the numbers drift out of date the moment someone forgets to log a change. A two-student capstone team set out to test whether a camera and a detection model could keep a running tally of phones in a storage cabinet with nobody touching a clipboard.

What We Built

We trained a vision model to recognize the two tracked phone models on the shelf from a live camera feed, then served the annotated video and a running tally to any browser on the network. Counts are appended to a records file every few seconds so the tally can be read back without re-running inference, and the whole pipeline runs on a small onboard computer right at the shelf. The detector scores strongly on its own validation split; the counting layer above it is a frame-by-frame heuristic with no object tracking, so a unit lifted out and put back is counted twice. That makes it a working proof of the approach rather than a production inventory system. Knowing exactly where the line falls is part of what shipping applied computer vision teaches you.

Results

• Runs at the shelf

  detection, video stream and tally all on one small onboard computer, no server involved

• Open in a browser

  the annotated feed and the running count reach any device on the network

• Logged continuously

  counts written to a records file every few seconds, so the tally reads back without re-running inference

• Proof of approach

  two trained classes and no object tracker, so a unit lifted out and put back is counted again

Techstack