Know exactly what's in every stream, every hour, every shift, every batch
Computer vision watches your recycling conveyor and reports composition continuously: how much PET, how much HDPE, how much contamination, by what time. Without robotic picking, without manual sampling.
Discuss a pilot →Quick facts
- Business size
- Recycling facilities, MRF operators, municipal sorting plants
- Timeline
- 4-6 weeks test, 2-3 months pilot, 3-6 months full rollout
- Budget range
- Pilot from €35k. Significantly cheaper than full sorting automation since no arm is required.
- Hardware
- Industrial cameras over the conveyor, Jetson-class edge compute, network for data export.
- Data needed
- Examples per material class, sometimes shared with sorting deployments if they exist.
- Evolution
-
- Genesis
- Custom-built
- Product
- Commodity
A vendor sells this result ready-made. We set it up and tune it to you.
What this scale means
Further right means more proven and cheaper. Further left means newer and riskier. Here is the test for each step.
- Commodity
- You could get the result yourself from a ready service, with almost no work. We rarely take these on.
- Product
- A vendor already sells this result turnkey, like shelf recognition from Trax or document reading from ABBYY. If one of them fits you, use it. You come to us when it does not: when it has to run on your own servers, cost less, or fit systems the product cannot reach.
- Custom-built
- Vendors sell only the parts. A tool like Tableau hands you charts, but the dashboards and metrics for your business still have to be built. That build is the work, and that is us.
- Genesis
- The approach exists but does not work reliably yet. You are betting on it maturing, so it costs more and carries more risk.
Expected outcomes
The Problem
Recycling operators know how much material they processed. They often do not know what was in it. The buyer specifies “PET bale, 90% pure or better”, the operator delivers a bale and hopes. The buyer’s lab samples the bale: sometimes it accepts, sometimes it rejects, sometimes it downgrades the price. The operator has no upstream visibility to catch problems before they ship.
The conventional approach is manual sampling: pulling a few kilos out of every batch for visual inspection. It is labor-intensive. It is statistically thin, since a few samples per shift can miss systematic problems. And it is lagged: you find out at 6pm that the morning shift was off-spec.
Computer vision (CV: cameras read by a model) watches the conveyor stream. This is the measurement-only sibling of robotic waste sorting. No arm, no picking, no facility rebuild, just continuous measurement of what moves down the line.
What the Solution Does
A camera (or several) mounted over your conveyor watches the stream and reports composition. Real-time, item-by-item, rolled up to whatever cadence matters for your decisions.
- Capture: cameras image the conveyor on a fast cadence (typically 1-4 frames per second for stream analytics).
- Detect: every visible object gets located and classified by material.
- Aggregate: composition statistics roll up per minute, hour, shift, or batch.
- Alert: contamination spikes or off-spec composition trigger notifications.
- Report: historical data feeds reports for buyer-side documentation and internal process improvement.
Where It Fits
This makes sense if you…
- Operate a recycling facility / MRF / sorting plant
- Sell graded output streams to buyers with purity specifications
- Currently rely on manual sampling for composition checks
- Want hourly visibility into what your sorting process (manual or automated) is actually producing
- Are not ready to invest in robotic sorting but want measurement first
This is probably not the right time if you…
- Process pre-sorted streams where composition is largely guaranteed upstream
- Have throughput so high that even high FPS won’t see most items (continuous monitoring is sampling at high frequency, still a sample)
- Need physical separation, beyond measurement (use automated waste sorting)
- Lack the setup to act on continuous data, with nobody using the hourly composition reports
Business Value
Continuous composition data. Manual sampling gives you 3-6 samples per shift. This gives you every-minute composition reports. That granularity is the difference between catching a contamination event during the shift and finding out after the shipment leaves.
Buyer-side documentation. Continuous monitoring gives you a record of every stream that goes out. “Bale 1247, produced 14:00-16:30, average 92% PET, 6% HDPE, 2% other” is a stronger story than “we believe this is PET”. Bale-level pricing and dispute defense improve.
Sorting process feedback. When manual sorters or robotic cells produce off-spec output, the system surfaces it within minutes. The sorting line gets coached or retuned before the problem builds up over a full shift.
Pilot before robotics. Many facilities weighing CV-guided sorting deploy stream monitoring first. It costs a fraction of the capex, it proves the underlying CV on your actual waste mix, and it builds the data foundation for later automation.
How It Works
The detection and classification stack is shared with robotic waste sorting. The difference here is no arm integration and a focus on aggregated reporting.
1. Camera setup
One or more cameras over the conveyor. Coverage depends on conveyor width and material density. We spec during the test phase.
2. Detection and classification
The system uses YOLO-family detection (a fast object detector) plus a fine-tuned material classifier. Output: a per-frame list of detected objects with material classes and confidence scores.
3. Stream aggregation
Per-frame detections feed a rolling aggregator: composition percentage by class, contamination rate, and throughput rate. Aggregates roll up to whatever cadence you need, by minute, hour, shift, or batch.
4. Alerting
Thresholds on contamination, off-spec composition, or unexpected throughput drops trigger alerts. These are configurable per stream, since different buyers tolerate different deviations.
5. Reporting and integration
Hourly, shift, and batch reports for buyer-side documentation. Data export to your existing BI tool or ERP. Optional integration with bale-level records for end-to-end traceability.
Stack
YOLO-family for detection. ResNet for material classification. Jetson-class edge compute. Datapipe runs the data pipeline into the warehouse. Metabase, Power BI, or Looker drives the operational dashboard.
What You Need to Make This Work
Data. Examples per material class (shared with sorting deployments if you have them). Initial corpus is typically 500-2000 images per class. In many cases we can train on existing facility footage.
Integrations. Camera mounting and power over the conveyor. Network for data export. Optional: link to bale records / ERP for traceability.
Hardware. Cameras and Jetson edge compute. No robotic arm. Much lower hardware budget than full sorting automation.
Team. A facility-ops lead. A quality-ops lead who will use the reports. An IT contact for network and data integration. Plan for around 15-25 hours during the pilot.
Implementation Roadmap
1. Test (4-6 weeks)
One conveyor, one stream. Install cameras, train a baseline model, and measure detection and classification accuracy on your specific waste mix. Validate composition reporting against manual sample counts. Output: a working camera setup, validated accuracy numbers, and recommendations for production rollout.
2. Pilot (2-3 months)
Production deployment. Integrate alert workflow. Build dashboards for ops and quality teams. Connect to bale records if in scope. Output: working production system, hourly composition data in use, documented bale-level reports.
3. Production (3-6 months)
Add more conveyors and more streams. Quarterly model retraining as the waste mix shifts seasonally. Continuous improvement through Datapipe.
Keep in Mind
Real limits:
- It measures, it does not sort. Stream monitoring tells you what is there. It does not separate. If you need physical separation, see automated waste sorting.
- Composition is a statistical estimate, never an exact count. The system sees what cameras can see, and overlapping or occluded items are partially hidden. Composition reports carry a statistical confidence, and we surface that openly.
- Material classes matter as much as the model. “Plastic” is one class. “PET, HDPE, PP” is three classes that need distinguishing. The model has to be trained for whatever granularity your buyer market requires.
- Throughput is conveyor-dependent. Fast conveyors mean less inference per item; slow conveyors mean better data but lower facility throughput. We tune during pilot.
- Lighting and conveyor mess matter. Dust, partial occlusion, grime on items all degrade classification. Camera placement and periodic recalibration are real maintenance items.
FAQ
Can this monitor multiple conveyors simultaneously?
Yes. Each conveyor gets its own camera and Jetson; backend processing aggregates across the facility.
How accurate is the composition reporting?
Typically ±5-10 percentage points by major category in good conditions. We validate against manual sampling during the test phase and report explicit numbers.
Can we use this as input to our buyer-side documentation?
Yes. The historical reports are designed to support bale-level documentation. Some buyers accept them directly; others use them as supporting evidence alongside their own lab samples.
What if we already have robotic sorting deployed?
Stream monitoring is complementary. It measures what the sorter is producing, surfaces drift, and supports continuous improvement of the sorting model.
How does this compare to chemical / spectroscopic stream analysis?
Different technology, different cost. Spectroscopic systems (NIR, X-ray) measure material chemistry directly and can tell apart polymers that look identical to a camera. They cost more and are more facility-specific. CV is the cheaper, broader-coverage option for general composition monitoring.
Ready to Discuss?
If you operate a recycling facility and want visibility into stream composition without the capex of full robotic sorting, this is a fast deployment. We will walk through your conveyor setup, your buyer specifications, and your sampling process, and tell you what to expect from a pilot.
Part of: Computer Vision ↗