Conveyor belt Wet-Dry detectioin and warning by image analysis
Budget: €12 – €18 EUR
I need a real-time vision/video routine that taps into our Unifi camera stream, inspects the texture of product (dewatered potato pulp) moving on a conveyor belt (or in a screw converyor), and decides whether that material is wet or dry. See images.
It is also possible to give more close up images of the product on the conveyor belt or during the overflow.
Once a decision is made the software must immediately do two things:
• push a flag we can read inside the PLC/SCADA layer so the process can react (speed change, diverter, etc.), and
• raise an operator alarm whenever the state flips or a confidence threshold is not met.
Key points you should know
• Source feed: live RTSP stream served by our Unifi NVR. No file uploads—the analysis has to run continuously.
• Core technique: texture analysis. We already have labelled clips showing clear examples of both states, which I will share for model training and validation.
• Output states: “Wet” or “Dry”, plus a confidence score.
• Performance expectation: sub-second latency and a configurable confidence threshold so we can tune false positives.
Deliverables
warning system that can trigger the operator.
More close up pictuers or camera positions can be deliverd if needed.
It is also possible to give more close up images of the product on the conveyor belt or during the overflow.
Once a decision is made the software must immediately do two things:
• push a flag we can read inside the PLC/SCADA layer so the process can react (speed change, diverter, etc.), and
• raise an operator alarm whenever the state flips or a confidence threshold is not met.
Key points you should know
• Source feed: live RTSP stream served by our Unifi NVR. No file uploads—the analysis has to run continuously.
• Core technique: texture analysis. We already have labelled clips showing clear examples of both states, which I will share for model training and validation.
• Output states: “Wet” or “Dry”, plus a confidence score.
• Performance expectation: sub-second latency and a configurable confidence threshold so we can tune false positives.
Deliverables
warning system that can trigger the operator.
More close up pictuers or camera positions can be deliverd if needed.