Conveyor sortation and parcel reading at a warehouse outbound

Starting point

At the warehouse outbound, parcels arrived on a single conveyor and staff read the label and carried each parcel to the right chute by hand. At peak hours two problems appeared together: line throughput dropped to the reading speed of the staff, and a parcel placed on the wrong chute only came back as a customer complaint.

The cause of a missort was invisible as well. There was no record — which parcel went where was not written down anywhere, only the total count at the end of a dispatch. At the end of a day you were left with “three parcels missing” and no idea at which chute or at what time those three went astray. Label quality was not uniform either: some parcels arrived with a creased barcode, others from a different supplier with a handwritten label.

The system built

The line was split into an identification zone and a sortation zone.

Identification zone. A fixed read tunnel was built over the conveyor: omnidirectional barcode readers plus a camera looking down. When the barcode cannot be decoded, OCR runs on the same image and tries to read the text field on the label. Why both methods: a barcode is fast and certain but has no defence against a damaged label; OCR is slow and probabilistic but generally reads what a human can read. The order matters — OCR only engages when the barcode has failed, otherwise every parcel carries unnecessary processing load.

Parcel tracking. An encoder was fitted to the conveyor drive, and every parcel is recorded together with the encoder value at the moment of identification. Its position up to the sortation point is calculated from that counter; because it is tied to distance rather than time, tracking does not drift when line speed changes or the line briefly stops. A counting photocell also sits on the conveyor: if the encoder-based expectation and the physical passage disagree, the line stops itself.

Recovery flow. A parcel where both the barcode and OCR fail is marked “unknown” and diverted to a separate recovery chute. There, an operator terminal shows the captured image of the parcel and the outcome of each read attempt; the operator enters the code by hand. The critical decision is this: an unknown parcel never goes to a default chute. A default chute is the most common source of silent missorts.

WMS integration. The mapping from parcel code to destination chute comes from the warehouse management system, and the sortation result is written back. The line does not stop when the link drops: the last mapping table received is held locally, result records are written to a queue and sent once the link returns.

For the control side see industrial automation, and for the reading side machine vision and camera systems.

The hard part

The unexpected difficulty was not reading but parcels touching each other. At a warehouse outbound, parcels do not arrive at regular intervals; two would enter the identification zone side by side or slightly overlapping. The reader then sees two barcodes while the system expects one parcel — the result being a code assigned to the wrong parcel. And the fault is silent: both parcels reach a valid chute, just not the correct one.

The first attempt was to separate them in software (deriving parcel boundaries from the image). With parcels of varying size, colour and surface finish this was not reliable; the seam between two dark parcels simply is not in the image.

The fix was to use mechanics and software together. A singulation conveyor with progressively increasing speed was added at the entrance of the identification zone, creating physical gaps between parcels. On the software side a gap check was added: if the encoder distance between two passages is below a threshold, the pair is not treated as read and goes straight into the recovery flow. In other words, when in doubt the system does not guess — it asks a human.

Measurement and acceptance

Three things were measured. Read rate: parcels decoded by barcode, recovered by OCR and dropped to the recovery chute are counted separately, with the ratios in a daily report. Sortation accuracy: over a defined shift, the system record was compared against an independent physical count at the chutes — a system verifying itself against its own record does not count as acceptance. Throughput: parcels sorted per unit time, on the same product mix.

FAT was run in the workshop: a set of real parcels was prepared and deliberately seeded with damaged labels, handwritten labels, labels applied upside down and parcels with no label at all. That the system rejected these and routed them to the recovery flow was verified case by case. SAT ran on real dispatch traffic with a signed protocol; the WMS link was deliberately cut to confirm the line kept running on its queue mechanism and that records transferred completely once the link came back. The numeric acceptance thresholds are written into the contract.