Intro & scope
Use cases where the primary input is images, video, point clouds or earth-observation rasters. Five patterns: VIS.INSPECT (quality, damage, NDT, survey), VIS.MONITOR (cameras observing people, traffic, animals, processes), VIS.IDENTIFY (recognise plates, faces, items, text, species), VIS.EARTH (satellite, aerial, SAR), VIS.MEDIA (search, moderation, anonymisation, video understanding). Boundaries: text-only document extraction belongs to DOC, sensor time series without imagery to PRED/ANOM, robot control to ROBO. Overlaps are resolved by input and decision: VIS-001/053/056/057 split by modality (RGB line cameras, X-ray/CT/UT, wafer maps, melt-pool/layer imagery); VIS-010 (generic intrusion) vs VIS-031 (natural-language VLM search over feeds); VIS-023 (claims photos) vs VIS-039 (field-work QA) vs VIS-037 (re-commerce grading); VIS-021 (generic EO) vs VIS-038/042/046/049/051/058/061 (domain EO). Regulatory note: remote biometric identification, emotion recognition and biometric categorisation are prohibited or high-risk under the EU AI Act (Art. 5, Annex III); GDPR DPIA is the default for any camera that captures people (rows with gdpr=yes). Industrial, non-personal inspection is typically minimal risk.
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| VIS-001 | Visual quality inspection & defect detection in manufacturing | VIS.INSPECT | mainstream | A | HW1 | minimal |
| VIS-002 | Assembly verification & completeness checks | VIS.INSPECT | mainstream | A | HW1 | minimal |
| VIS-003 | Infrastructure inspection with drones (bridges, power lines, wind turbines) | VIS.INSPECT | early-adopter | A | HW3 | depends |
| VIS-004 | Road damage & public asset condition monitoring | VIS.INSPECT | early-adopter | A | HW1 | minimal |
| VIS-005 | Plant disease & crop health detection | VIS.INSPECT | early-adopter | A | HW1 | minimal |
| VIS-006 | Food quality grading & sorting | VIS.INSPECT | mainstream | A | HW1 | minimal |
| VIS-007 | Workplace safety & PPE compliance monitoring | VIS.MONITOR | early-adopter | A | HW1 | depends |
| VIS-008 | People counting & footfall analytics | VIS.MONITOR | mainstream | A | HW1 | minimal |
| VIS-009 | Retail shelf monitoring & planogram compliance | VIS.MONITOR | early-adopter | A | HW2 | minimal |
| VIS-010 | Smart video surveillance & intrusion detection | VIS.MONITOR | mainstream | A | HW1 | depends |
| VIS-011 | License plate recognition & parking management | VIS.IDENTIFY | mainstream | A | HW1 | minimal |
| VIS-012 | Traffic monitoring & analytics | VIS.MONITOR | mainstream | A | HW1 | depends |
| VIS-013 | Wildfire & environmental monitoring from cameras | VIS.MONITOR | early-adopter | A | HW2 | minimal |
| VIS-014 | Livestock & animal welfare monitoring | VIS.MONITOR | early-adopter | A | HW1 | minimal |
| VIS-015 | Sports analytics & player tracking | VIS.MONITOR | mainstream | A | HW2 | minimal |
| VIS-016 | Face recognition & biometric access control | VIS.IDENTIFY | mainstream | A | HW1 | depends |
| VIS-017 | Visual product recognition & checkout-free retail | VIS.IDENTIFY | early-adopter | A | HW3 | depends |
| VIS-018 | Industrial OCR: labels, serial numbers, package tracking | VIS.IDENTIFY | mainstream | A | HW1 | minimal |
| VIS-019 | Warehouse inventory counting with cameras & drones | VIS.IDENTIFY | early-adopter | A | HW1 | minimal |
| VIS-020 | Waste sorting & recycling recognition | VIS.IDENTIFY | early-adopter | A | HW1 | minimal |
| VIS-021 | Satellite & aerial imagery analysis (land use, change detection) | VIS.EARTH | mainstream | B | HW3 | minimal |
| VIS-022 | Visual search (search by image) | VIS.MEDIA | mainstream | B | HW2 | minimal |
| VIS-023 | Damage assessment from photos (vehicle, property claims) | VIS.INSPECT | mainstream | A | HW2 | minimal |
| VIS-024 | Pose & gesture analysis (ergonomics, fitness, rehab) | VIS.MONITOR | early-adopter | B | HW1 | depends |
| VIS-025 | Video understanding, tagging & highlight detection | VIS.MEDIA | early-adopter | B | HW3 | limited |
| VIS-026 | Visual measurement, counting & volume estimation | VIS.INSPECT | mainstream | B | HW1 | minimal |
| VIS-027 | Image & video content moderation (nudity, violence, CSAM, terrorist content) | VIS.MEDIA | mainstream | A | HW2 | minimal |
| VIS-028 | Driver & in-cabin monitoring (drowsiness, distraction, fleet safety dashcams) | VIS.MONITOR | mainstream | A | HW1 | depends |
| VIS-029 | Facial age estimation & age assurance | VIS.IDENTIFY | early-adopter | A | HW1 | depends |
| VIS-030 | Image & video anonymisation (face & licence-plate blurring) | VIS.MEDIA | mainstream | A | HW2 | minimal |
| VIS-031 | Natural-language video search, alerts & summaries over camera feeds (VLM video agents) | VIS.MONITOR | early-adopter | A | HW4 | depends |
| VIS-032 | Species identification & camera-trap biodiversity monitoring | VIS.IDENTIFY | mainstream | C | HW2 | minimal |
| VIS-033 | Food waste tracking in commercial kitchens (camera + scale smart bins) | VIS.MONITOR | mainstream | A | HW1 | minimal |
| VIS-034 | Manual work cycle-time & line-balancing analytics from video | VIS.MONITOR | early-adopter | A | HW1 | depends |
| VIS-035 | Visual brand & logo detection for sponsorship valuation & visual social listening | VIS.MEDIA | mainstream | A | HW2 | minimal |
| VIS-036 | Creative pre-testing with predicted visual attention & ad-effectiveness scoring | VIS.MEDIA | mainstream | A | HW2 | minimal |
| VIS-037 | Returns & second-hand item condition grading (re-commerce) | VIS.INSPECT | early-adopter | B | HW2 | minimal |
| VIS-038 | Earth-observation compliance checks for public authorities (CAP area monitoring, undeclared construction) | VIS.EARTH | early-adopter | B | HW3 | depends |
| VIS-039 | Photo-based field work verification (installation QA at point of capture) | VIS.INSPECT | early-adopter | A | HW2 | depends |
| VIS-040 | Scan-to-BIM & digital site survey (as-built models and floor plans from point clouds & phone scans) | VIS.INSPECT | early-adopter | A | HW3 | minimal |
| VIS-042 | Methane & greenhouse-gas plume detection from satellites | VIS.EARTH | early-adopter | A | HW3 | minimal |
| VIS-043 | Sewer & pipe CCTV inspection with automatic damage coding (EN 13508-2) | VIS.INSPECT | early-adopter | A | HW2 | minimal |
| VIS-044 | Drone detection & counter-UAS for airports & critical infrastructure | VIS.MONITOR | early-adopter | A | HW2 | depends |
| VIS-045 | AI-assisted X-ray & CT security screening (baggage, parcels, cargo) | VIS.INSPECT | mainstream | A | HW1 | depends |
| VIS-046 | Automated PV & heat-pump system design and quoting from aerial imagery | VIS.EARTH | mainstream | A | HW2 | minimal |
| VIS-047 | Self-checkout loss prevention (missed-scan & ticket-switch detection) | VIS.MONITOR | early-adopter | A | HW1 | minimal |
| VIS-048 | Aircraft turnaround & logistics process milestone monitoring | VIS.MONITOR | early-adopter | A | HW2 | depends |
| VIS-049 | Deforestation-free supply-chain verification (EUDR) & forest carbon MRV | VIS.EARTH | early-adopter | A | HW3 | minimal |
| VIS-050 | Digital forensics & seized-media triage (incl. CSAM victim identification) | VIS.IDENTIFY | early-adopter | A | HW3 | high |
| VIS-051 | Geospatial intelligence & ISR imagery analysis (defence & security) | VIS.EARTH | early-adopter | A | HW3 | depends |
| VIS-053 | Automated NDT evaluation (industrial X-ray, CT, ultrasonic) for castings & welds | VIS.INSPECT | early-adopter | A | HW3 | depends |
| VIS-054 | AI traffic-offence enforcement cameras (phone use at the wheel, seatbelt violations) | VIS.MONITOR | early-adopter | A | HW1 | depends |
| VIS-055 | Automated AI camera production for live sports & event streaming | VIS.MEDIA | mainstream | A | HW2 | minimal |
| VIS-056 | Semiconductor wafer-map defect classification & fab yield analytics | VIS.INSPECT | mainstream | A | HW2 | minimal |
| VIS-057 | In-situ process monitoring & layer-wise defect detection in metal additive manufacturing | VIS.INSPECT | early-adopter | A | HW2 | minimal |
| VIS-058 | Rapid post-disaster building damage mapping from satellite & drone imagery | VIS.EARTH | early-adopter | A | HW3 | depends |
| VIS-059 | Remote electronic monitoring (REM) of fishing vessels with AI review of catch & discard footage | VIS.MONITOR | early-adopter | A | HW2 | minimal |
| VIS-060 | Camera-based bird collision avoidance & shutdown control for wind turbines | VIS.MONITOR | early-adopter | A | HW1 | depends |
| VIS-061 | InSAR ground-motion monitoring for infrastructure, mining & subsidence | VIS.EARTH | early-adopter | A | HW2 | minimal |
Shared stack
Common building blocks (same capability tags): object-detection/segmentation/image-classification (detector backbone), tracking, edge-runtime, vlm (open-vocabulary, triage, captioning), anomaly-detection (few-defect lines), ocr-docparse (labels/plates), remote-sensing, 3d-vision.
HW0/HW1 - edge box or workstation GPU (up to ~24 GB, 1-16 camera streams) - Ingest: RTSP/GigE Vision via GStreamer or Frigate/DeepStream; MQTT or OPC UA to PLC/MES. - Models: YOLO-family or RT-DETR detectors, SAM-class segmenters, PatchCore/anomalib for anomaly detection, PaddleOCR/EasyOCR; export to ONNX/TensorRT/OpenVINO. - Runtime: NVIDIA Jetson Orin or a single RTX-class GPU, Triton or ONNX Runtime; on-device blurring before storage (VIS-030 module reused by VIS-007/008/010/011/028). - Ops: CVAT or Label Studio labelling, MLflow registry, Prometheus/Grafana drift and latency, S3-compatible (MinIO) frame store with short retention. - Fits: VIS-001/002/006/008/011/016/018/026/028/029/033/034/045/047/054/060.
HW2 - single server, 1-2 GPUs (48-96 GB VRAM) - Everything in HW1 plus a self-hosted VLM (Qwen-VL / InternVL class, vLLM) for open-vocabulary checks and report drafting, an embedding model (SigLIP/CLIP) with pgvector or Qdrant for visual search and re-identification-free matching. - Pipelines: Dagster/Airflow batch inference, Postgres + object store, Keycloak SSO, review UI with human-in-the-loop sign-off. - Fits: VIS-009/013/015/022/023/027/030/032/035/036/037/039/043/044/046/048/055/056/057/059/061.
HW3 - GPU cluster or EU cloud GPU, tiled/large-volume imagery - Kubernetes with GPU nodes (Ray or Kubeflow), STAC catalogue over Sentinel-1/2 (Copernicus/CDSE), COG/Zarr storage, TorchGeo and EO foundation models (Prithvi, Clay), photogrammetry (COLMAP/OpenDroneMap, Open3D) for drone and scan-to-BIM work. - Strict EU data residency (EU sovereign cloud or on-premises), audit logging, model cards for AI Act documentation. - Fits: VIS-003/017/021/025/038/040/042/049/050/051/053/058.
HW4 - multi-node serving for large video-language models - vLLM/SGLang multi-GPU serving of a 70B-class VLM, streaming keyframe sampler, vector store over clip embeddings, agent layer with guardrails and role-based retrieval; used by VIS-031 and reusable for VIS-025 and KNOW-016 video retrieval.
Cross-family reuse: embedding + vector DB layer shared with KNOW (KNOW-005/016), OCR/layout stack with DOC (DOC-001/006), time-series/anomaly services with PRED/ANOM, robotic cells with ROBO-001, EO/geodata with DATA-011/PRED-027. One shared MLflow, labelling and monitoring platform serves all tiers; treat people-facing deployments (VIS-007/010/016/028/047/054) as a separate namespace with DPIA, retention limits and logging.