Warehouse inventory counting with cameras & drones

Annual stocktakes and manual cycle counts stop operations, cost labour and still leave location errors undetected. Camera-equipped drones, forklift-mounted cameras and fixed cameras read pallet labels, barcodes and shelf contents continuously, compare them with the warehouse management system (WMS) and flag only the discrepancies. The result is higher inventory accuracy, fewer stock-outs and counts done at night or between shifts without halting picking.

Problem & value

• Today: staff walk aisles with handheld scanners, or the site closes for a wholesale stocktake. High racking (above 8-10 m) needs lift trucks or aerial work platforms, which is slow and a safety risk. Misplaced pallets, wrong SKUs and phantom stock surface late, often as missed orders.
• AI value: frequent, evidence-backed counts (photo per location), safer work at height, labour moved from walking to resolving exceptions, and a continuously updated view of occupancy.
• Evidence: vendors such as Verity, Gather AI and Corvus publish customer case studies on count speed and accuracy; treat these as vendor claims and validate with a pilot on your own racks [1][2][3]. Regulators do not publish pooled benefit figures for this use case.

Variants

• Autonomous indoor drones scanning racks at night (high-bay racking, large 3PL sites).
• Forklift- or AMR-mounted cameras that scan locations as vehicles go about normal work.
• Fixed cameras / gantry or portal scanners at dock doors and conveyor lines for inbound and outbound verification.
• Handheld / phone-based visual counting of loose goods and shelf stock (see VIS-026, VIS-009).
• RFID-augmented counting where tagged items are read in bulk; cameras cover untagged pallets.
• Outdoor yards and bulk stock (timber, coal, scrap) using drone photogrammetry for volume (VIS-026).

How it works

1. mission plan / vehicle route (rack map, location list)
2. image & video capture (drone, forklift camera, fixed camera) + localisation (visual SLAM / UWB / lidar)
3. detection of pallets, cartons and labels (object detection, segmentation)
4. barcode / QR / label OCR, SKU and lot reading (OCR, VLM for damaged labels)
5. location assignment (pose + rack map
6. bin ID)
7. reconciliation with WMS (expected vs observed)
8. exception queue (human re-check with photo)
9. cycle-count task / WMS adjustment

• Localisation is the hard part indoors without GPS: visual-inertial SLAM, lidar or UWB anchors tie each image to a bin.
• Reading: 1D/2D codes are decoded with classical or ML decoders; label text (SKU, lot, expiry) with OCR or a VLM, which handles motion blur and angled shots better than template OCR (see VIS-018).
• Counting units inside a pallet (cartons per layer) uses detection or a VLM; accuracy drops for stretch-wrapped, mixed or occluded pallets.
• Human-in-the-loop: discrepancies go to an inventory controller with the photo; WMS stock adjustments above a value threshold require approval. Drones and AI do not post stock changes alone.

Data preparation

Counting only creates value when it can be compared with the book stock, so the key data is the WMS/ERP stock list: article master, storage locations, units of measure and pallet or case sizes. In an SMB this lives in the ERP, in Excel location sheets and sometimes only in the heads of warehouse staff. Typical gaps are missing or inconsistent location codes, no barcodes or labels on racks, mixed packaging units, duplicate articles and no photos of how stock actually looks on the shelf (stacked, wrapped, partially hidden).

Preparation steps: clean the location structure and label racks with readable codes or barcodes, export and de-duplicate the article master (see PREP-005 and PREP-011), define units and tolerance rules for what counts as a discrepancy, and run one manual count in parallel to create a verified test set (PREP-018). Tools: the ERP export, a labelling tool and the vendor's site-mapping app. Data and site preparation is roughly 25-35 percent of effort; the physical labelling of racks is often the biggest item.

Implementation paths

Which path for whom

• SMB: Path A, because packaged camera, scanner-drone or shelf-scan services exist for single sites and avoid building a vision pipeline.
• Enterprise: Path B, because counts must reconcile with WMS/ERP stock across many sites and an own pipeline avoids per-site licence costs.
• Developer: Path C, because open detection and OCR models running on edge devices allow custom counting products without cloud dependency.

Path A — Buy (recommended for most sites)

• Drone-based: Verity (Swiss, non-EU), Gather AI (US, drones plus forklift cameras), Corvus Robotics (US), Eyesee (Hardis Group, FR/EU), Dronegy (FR/EU), Doks Innovation (Croatian/EU, Innovation drones).
• Vehicle- and fixed-camera: Gather AI forklift modules, Zebra (scan and machine vision), Scandit (SDK for smart-device scanning, Swiss/US), Cognex and Datalogic (fixed readers, EU/US).
• Check: WMS connector (SAP EWM, Manhattan, Blue Yonder, Körber, Oracle), rack-height range, safety certification for operation near people, data location and retention of images.
• Choose A when: a site has several thousand pallet positions, racking above 6 m, or an inventory-accuracy problem that justifies a pilot.

Path B — Build on cloud APIs

• Reading and counting: multimodal models via EU-region endpoints (Gemini on Vertex AI, Azure OpenAI, Claude via cloud providers, Mistral) for label and pallet understanding; Azure AI Document Intelligence or Google Document AI for label text; AWS Rekognition text detection.
• Orchestration: event pipelines (MQTT/Kafka) from capture devices, n8n or custom services, and WMS APIs for expected stock.
• Choose B when: you already capture images (forklift tablets, cameras) and need custom logic.

Path C — Self-host (open source)

• Detection: Ultralytics YOLO (note AGPL licence) or permissively licensed RT-DETR/D-FINE variants; Grounding DINO / SAM family for open-vocabulary and segmentation; supervision/ByteTrack for tracking.
• Codes and text: ZXing-cpp, zbar, PaddleOCR; open VLMs (Qwen-VL family) via vLLM for hard labels.
• Mapping: ROS 2 with RTAB-Map or ORB-SLAM3 for localisation; NVIDIA Isaac ROS where Jetson is used.
• Hardware: HW1 (Jetson Orin class on vehicle/drone, CPU server) for detection and decoding; HW2 for a VLM serving several sites. Drone hardware itself is bought, not self-built.

Path D — Custom training

Worth it for unusual label layouts, wrapped or mixed pallets and site-specific rack geometry. A few thousand labelled images per site type are usually needed; operator corrections from the exception queue provide training data. Retrain per site cheaply by fine-tuning a small detector.

State of the art (2026)

• Mainstream building blocks, early-adopter systems: code reading and pallet detection are mature; end-to-end autonomous scanning of mixed warehouses is in production at leading 3PLs and retailers but still needs per-site setup and mapping.
• VLMs improved reading of damaged, curved or low-light labels and counting visible cartons, but can hallucinate plausible SKUs: always validate reads against check digits, WMS masters and GS1 formats (GTIN, SSCC) [4].
• Fusion of RFID, camera and WMS data is becoming the standard for high accuracy.
• Limits: items hidden behind others, inside closed cartons or in the second pallet depth cannot be counted visually; reflective wrap, low light, dust and motion blur reduce read rates; drone battery life limits coverage per mission; indoor flight near people needs strict safety rules.

Shared resources & combinations

• Same stack (object-detection + ocr-docparse + edge-runtime): VIS-018 industrial OCR, VIS-009 shelf monitoring, VIS-026 counting and volume estimation, VIS-017 product recognition.
• Drone and mapping platform: shared with VIS-003 (drone inspection) and ROBO-006 (autonomous drones).
• Downstream: OPT-005 replenishment and OPT-006 slotting consume accurate stock and occupancy data; VIS-007 can reuse the same cameras for safety if governance allows (separate purpose).
• Logistics documents: DOC-006 for matching delivery notes to scanned pallets.

Effort, cost & KPIs

• Path A: pilot in 2-3 months for one aisle block, rollout over a few quarters; pricing is typically subscription per site or per drone/vehicle (obtain quotes; figures vary widely).
• Path B: 3-6 months; mainly capture hardware, localisation and WMS integration.
• Path C/D: 6+ months, specialised robotics and CV team.
• KPIs: inventory record accuracy (location and quantity), read rate per location, locations scanned per hour, share of discrepancies confirmed by humans (precision), stock-outs and mis-picks, labour hours per count, safety incidents at height.

People & change

Start with management for scope and budget and the warehouse lead as process owner; staff who count today must be part of the design because counting rules and exceptions live in their experience. IT and IT security cover network, drone or camera access and data flows; the data owner of ERP stock data signs off the reconciliation rules. Inventory of internal stock is business data, but cameras and drones also capture employees, so the works council has co-determination under §87(1) no. 6 BetrVG (objective suitability suffices according to the BAG) and must be informed under §90(1) no. 3. Agree in a works agreement that images are used only for stock counting, are not used for performance review and are deleted or blurred after a short period. If staff are identifiable, involve the DPO, check the need for a DPIA and sign an AVV with cloud vendors.

Support AI literacy per Art. 4 with a short training on what the counts mean and when to trust them. Choose one aisle or one site as pilot and run it alongside manual counts. Common concerns are surveillance and job loss; address them with transparency about flight and camera zones and by moving staff to exception checks. See GUIDE-004, GUIDE-009 and GUIDE-013.

Risks, regulation & compliance

• EU AI Act: minimal risk for stock counting. It becomes sensitive if cameras are also used to evaluate workers (productivity monitoring, emotion inference). Emotion recognition in the workplace is prohibited under Art. 5(1)(f); worker-management AI can be high-risk under Annex III point 4 [5]. Keep counting systems purpose-limited.
• GDPR: images will capture employees and drivers. Use privacy-by-design: purpose limitation, blurring or cropping people (see VIS-030), short retention of raw video, a DPIA where continuous monitoring is likely, and works-council co-determination in Germany (BetrVG §87(1) no. 6) [6].
• Aviation and safety: drones indoors are generally outside EASA open-category scope when fully enclosed, but check national rules, insurance and the machinery/workplace safety assessment [7]. Outdoor yard flights fall under Regulation (EU) 2019/947.
• Data and cybersecurity: inventory data is commercially sensitive; secure drone and camera networks, signed firmware and EU hosting where required.
• Mitigations: human approval of stock adjustments, photo evidence per location, shadow-mode comparison with manual counts before replacing audits, auditor acceptance of the method (accounting stocktake rules such as HGB §240/§241 permit permanent inventory procedures under conditions).

Getting started

1. Measure the baseline: inventory accuracy, count cost, count frequency and top mismatch causes.
2. Pick one rack zone with barcode-labelled locations and run a 4-8 week pilot with 1-2 vendors.
3. Compare results against manual cycle counts at the same locations (read rate, precision, speed).
4. Agree GDPR, works-council and safety measures before flights or camera rollout.
5. Integrate exceptions into the WMS task queue, then expand zone by zone.

Sources

1. Verity — autonomous inventory drones — https://verity.net
2. Gather AI — autonomous inventory monitoring — https://www.gather.ai
3. Corvus Robotics — Corvus One inventory drone — https://www.corvus-robotics.com
4. GS1 — GTIN and SSCC identification standards — https://www.gs1.org/standards/id-keys
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. EDPB — Guidelines 3/2019 on processing of personal data through video devices — https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-32019-processing-personal-data-through-video_en
7. EASA — Drones and air mobility, rules for drone operations — https://www.easa.europa.eu/en/domains/civil-drones
8. Ultralytics YOLO documentation — https://docs.ultralytics.com
9. ROS 2 documentation — https://docs.ros.org