Skip to content
Onigiri Tech

Computer Vision / AI

Visual Inspection & Tracking System

A camera-based pipeline that detects, counts and inspects items moving through a loading area, and turns each event into searchable data.

  • Object detection
  • Multi-object tracking
  • Label OCR
  • Edge inference
  • Review interface
  • Event alerts

Context

A logistics operation relying on manual counts and spot checks at loading bays, with existing cameras used only for occasional review.

The challenge

Counting and inspection happened on paper, so discrepancies were discovered days later.

Footage existed but was never searchable — finding one event meant scrubbing through hours of video.

Approach

  1. 01

    Feasibility on real footage

    Camera angles, lighting and image quality evaluated before any model work began.

  2. 02

    Detection, tracking and OCR

    Items detected and tracked across frames, with label text read and matched against expected shipments.

  3. 03

    Edge-first deployment

    Inference runs on-site; only events, crops and metadata are sent to the central system.

  4. 04

    Humans review the uncertain cases

    Low-confidence detections are queued for review, and corrections feed future retraining.

What we built

  • Inference pipeline
  • Event store
  • Review queue
  • Search by label or time
  • Alerts
  • Operations dashboard
  • Python
  • PyTorch
  • OpenCV
  • ONNX Runtime
  • PostgreSQL
  • Next.js

Outcome

  • Counts recorded automatically per bay and per shift
  • Every event searchable by time, bay or label text
  • Uncertain detections routed to people instead of silently accepted

Next case study

Client Service Portal & Mobile App

Complex problem? Good.

Have a software problem worth solving?

Whether you're starting with an idea, replacing an existing system, or trying to automate an operation that has become too complicated, let's talk.