Real-Time Produce Stock Counter
Budget: $1,500 – $3,000 AUD
I want to replace manual stock tallies in our Sydney fresh-produce warehouse with a computer-vision system that works in real time. The cameras will sit above the loading docks, stream a live video feed, and your model should:
• Detect every incoming or outgoing box
• Identify whether the box contains tomatoes, cucumbers, or eggplants
• Update inflow / outflow counts automatically in our database (PostgreSQL)
I’ll supply sample footage, camera specs, and a schema for the existing warehouse database. I’m expecting a functional prototype that includes:
1. A trained detection model (YOLO, Detectron2, or similar) with clear instructions for retraining as product packaging changes.
2. An inference service that ingests the live RTSP stream, performs detection on the edge or on an on-prem GPU, and pushes updates via a database sync routine.
3. A lightweight web dashboard (React or simple Flask template is fine) that shows current on-hand counts, recent movements, and a searchable activity log.
4. Deployment scripts or Dockerfiles so we can replicate the setup on additional loading bays.
5. Brief documentation covering installation, configuration, and how to add new produce SKUs.
Accuracy and speed matter: detections should stay above 95 % precision at a minimum 10 fps on a 1080p stream. Keep the code modular so maintenance staff can swap cameras or tweak thresholds without touching the model internals.
If this scope sounds comfortable and you have proven experience with real-time computer vision systems, let’s talk timelines and milestone testing on our live cameras.
• Detect every incoming or outgoing box
• Identify whether the box contains tomatoes, cucumbers, or eggplants
• Update inflow / outflow counts automatically in our database (PostgreSQL)
I’ll supply sample footage, camera specs, and a schema for the existing warehouse database. I’m expecting a functional prototype that includes:
1. A trained detection model (YOLO, Detectron2, or similar) with clear instructions for retraining as product packaging changes.
2. An inference service that ingests the live RTSP stream, performs detection on the edge or on an on-prem GPU, and pushes updates via a database sync routine.
3. A lightweight web dashboard (React or simple Flask template is fine) that shows current on-hand counts, recent movements, and a searchable activity log.
4. Deployment scripts or Dockerfiles so we can replicate the setup on additional loading bays.
5. Brief documentation covering installation, configuration, and how to add new produce SKUs.
Accuracy and speed matter: detections should stay above 95 % precision at a minimum 10 fps on a 1080p stream. Keep the code modular so maintenance staff can swap cameras or tweak thresholds without touching the model internals.
If this scope sounds comfortable and you have proven experience with real-time computer vision systems, let’s talk timelines and milestone testing on our live cameras.