Vision Index

Phase 5 audit · 2026-09-17

Market study

Compare edge-runnable computer vision models and commercial solutions for camera-agnostic compliance — smoking, PPE/gloves, fire/smoke, and safety — so a researcher can pick a path and prove it on site.

Live catalog: 105 entries (98 active) · 30 devices.

105
Entries
98
Active
7
Eliminated
30
Devices
3
Verified

Catalog · Hardware

Scope

Violations

  • Smoking / cigarette
  • PPE (helmet, vest, gloves)
  • Fire & smoke
  • Health (handwash, hairnet)
  • Fall & behavior

Mandate

  • On-prem / edge inference (no cloud-only)
  • Camera-agnostic RTSP/ONVIF ingest
  • 1080p ingest · 15–30 FPS target
  • 640×640 detector input (Coral ≤512px)
  • Site benchmark before any SLA

How to use

  1. Pick a use case below or on the home page.
  2. Open the filtered catalog and compare at most three entries.
  3. Read smoking and glove gaps before you write an .
  4. Match cameras to a hardware tier, then run an on-site .

Verdict

General detection is solved; compliance is not. No verified smoking mAP on hold-out CCTV. Glove recall is the weak PPE class. Commercial smoking+PPE POC: Lumeo → Dragonfruit → FuweeVision. DIY: YOLO11 + SH17 + merged smoking sets. No production SLA without on-site validation.

Phase 4 said no major OEM lists smoking and ranked Coram/Spot first. The Phase 5 audit found documented smoking/vaping modules (accuracy still unverified) and moved Lumeo to #1 for smoking+PPE POC.

Key findings

  • General detection is solved; compliance is not. COCO backbones (YOLOv8/11, RT-DETR, RF-DETR) reach 37–54 mAP50-95 with fine-tuning. Pretrained PPE and fire/smoke weights exist. Smoking has no verified pretrained SOTA.
  • Gloves are the PPE weak link. Verified floor: recall 0.555 (DarthRegicid). itsadityabaniya claims Gloves mAP50 ~0.69 (unverified). SH17 released 17-class weights. Bench all three before an SLA.
  • Agnostic edge boxes beat camera OEM analytics for mixed RTSP. Hikvision, Dahua, Axis score low agnostic fit. Lumeo, Coram, Spot AI, and Jetson DIY score high.
  • Hardware picks model size, not architecture. Orin Nano: YOLOv8n/s, ≤4 cameras. Orin NX: YOLOv8m, 4–12 cameras. RK3588: YOLOv8n real-time; YOLOv8m ~12 FPS. Coral: ≤512px input.
  • NL/VLM is a supplement, not the default. Ambient, Coram Custom Detection, Spot Iris are experimental for cigarette compliance. Prefer deterministic pipelines (Lumeo smoking classifier, Dragonfruit vaping) until a side-by-side bench.

Method

Collect everything, then filter. Every accuracy and FPS claim is tagged. Phase 5 searched independently against the catalog, added nine missed solutions, and confirmed all seven eliminations.

Phase 1
Discovery
Phase 2
Verification
Phase 3
Comparison
Phase 4
Final report
Phase 5
Audit

Open-source 59 · commercial 46 · eliminated 7. Phase 4 started at 96 entries; the audit expanded the catalog to 105 (98 active).

Detection coverage

Helmet / vest Strong

Verified + partial weights

Gloves Weak

Recall 0.555 verified floor

Smoking Documented

Lumeo / Dragonfruit — accuracy unverified

Fire / smoke Good

imnuman mAP50 0.872

Handwash / hairnet Partial

Restaurant-specific datasets

Fall Partial

Pose + heuristics; site calibration

PathPPEGlovesSmokingFireAgnostic
YOLO11 DIY + fine-tune✓△ trainTrain required✓High
DarthRegicid PPE✓ verified0.555 recall✗✗High
Lumeo pipelines✓ documented✓ gloves pageUnverified✓High
Coram Point (Jul 2025)Native PPEIncl. glovesNot documentedPartialHigh
Dahua WizSense OEMGeneric PPENot broken out✗✗Low
Scylla smoke module✗✗≠ cigarette✓High

Decision flow

Choose the deployment path first, then the model.

ML / engineering team?

Yes — open-source DIY (YOLO11 + fine-tune).

No — commercial edge appliance (Lumeo / Coram / Spot).

Then constrain

Budget → RK3588 + YOLOv8n.

Smoking + PPE → Lumeo POC.

Enterprise SLA → Ipsotek VISuite.

OEM fleet → Dahua WizSense PPE. Mixed cameras → agnostic edge box.

If #1 is…PickAvoid
Lowest TCORK3588 + YOLOv8nAmbient.ai, Ipsotek enterprise
Highest accuracyRF-DETR or YOLO11 fine-tunedCoral TPU, vendor_claim rows
Camera agnosticLumeo, Coram, Spot, Jetson DIYHikvision, Dahua, Verkada closed
Smoking complianceLumeo → Dragonfruit; DIY UiTM mergeOEM “smoke” modules; NL as default
Glove complianceTriple OSS bench + HexmonCommercial PPE without a glove SLA
No ML teamLumeo, Coram (PPE), DragonfruitDIY YOLO11, MMDetection
Factory multi-camOrin NX + DeepStream + YOLO11Orin Nano beyond 4 cameras

Use-case shortlists

Phase 5 POC order. All require site validation and written detection rates.

Smoking + PPE POC
Lumeo → Dragonfruit → FuweeVision

Only top-tier vendors with documented smoking and glove pipelines. Accuracy still unverified — require written rates.

Fastest PPE / SOP
Coram Point or Spot AI IVR

Native PPE including gloves (Coram Jul 2025). Smoking is not documented — do not use this path for cigarette compliance.

DIY accuracy control
YOLO11 + SH17 + UiTM / fire-exit

Triple glove bench; merge UiTM (1940) and fire-exit CCTV (8124) for smoking train. AGPL on Ultralytics — legal review for closed SaaS.

Lowest cost per site
RK3588 + YOLOv8n

~53 FPS on budget ARM. YOLOv8m is marginal at ~12 FPS. Coral is cheaper but capped at 320–512px.

Factory multi-cam
Orin NX + DeepStream + YOLO11

4–12 cameras at 15–30 FPS. Mature production stack. AGX Orin if you need 12+ streams.

APAC manufacturing
Staqu JARVIS or XINHUO

Explicit gloves; industrial smoking+PPE pages. Added in the Phase 5 audit.

Agnostic edge box

Phase 4: Coram · Spot · Lumeo. Phase 5: Lumeo first for smoking+PPE.

1Lumeo Video Analytics Platform — Only top-tier pick with published smoking + glove pipelines
2Coram Point AI NVR — Native PPE incl. gloves; smoking unconfirmed
3Spot AI IVR + AI Agents — PPE/SOP strong; Iris NL for custom; smoking not documented
+Dragonfruit AI Frontier — Vaping/smoking on safety suite (~$1k/site/yr)
+FuweeVision Edge AI Algorithm Box — Industrial smoking+PPE edge boxes

Open-source DIY

Added SH17 weights and a second glove baseline; smoking train set is a merge.

1Ultralytics YOLO11 — Default backbone — tooling and edge export
2SH17 Dataset + YOLOv8/v9/v10 PPE Weights — 17-class PPE, released checkpoints
3DarthRegicid1/YOLOv5_PPE-Detection — Bench both for gloves; 0.555 is not a ceiling
4Cigarette Detection Dataset V2 (Roboflow/UiTM) — Merge 1940 + 8124-image smoking sets

Commercial turnkey

Lumeo and Dragonfruit elevated; Staqu added for APAC.

1Lumeo Video Analytics Platform — No-code smoking + gloves
2Dragonfruit AI Frontier — Vaping/smoking listed on safety suite
3IntelliVision AI Video Analytics — OEM/enterprise PPE bundles
+Staqu JARVIS AI Video Analytics — APAC manufacturing; explicit gloves

Ten solution groups

Counts are live from the catalog. Open a group to browse entries.

General detection

18

COCO backbones for custom fine-tune. None ship smoking or glove classes out of the box.

YOLO11 for new work. RF-DETR if Apache is required. YOLOv8 for the verified COCO baseline.

Open-vocab / VLM

3

Text-prompt detection without a dedicated class. CCTV compliance accuracy is unvalidated.

Phase 5 experiment, not the production default. Pair with a fine-tuned YOLO for gloves and smoking.

PPE weights

13

Pretrained PPE checkpoints. Glove recall is the weak class — bench three baselines before SLA.

Start at DarthRegicid. Bench itsadityabaniya and SH17. Add Hexmon NO-Gloves. Target ≥0.70 recall on site.

Smoking, fire & safety

9

Fire/smoke has usable weights. Smoking has datasets and papers — no verified pretrained SOTA.

Fire: imnuman/fire-detection-yolo. Smoking: train YOLO11n on UiTM (1940) + fire-exit (8124).

Behavior & pose

5

Fall, posture, and pose pipelines. Complementary to compliance detectors, not a substitute.

MediaPipe for light pose. YOLO Pose + heuristics when it shares the violation pipeline.

Runtimes & export

8

Toolchains, not detectors. Pick after the model — TensorRT, RKNN, Hailo, OpenVINO, NCNN.

DeepStream on Jetson. RKNN on Rockchip. MMDeploy as the vendor-neutral export hub.

Camera OEM analytics

13

Bundled camera/NVR analytics. Fast on a locked fleet, poor camera-agnostic fit.

Use only on a locked camera fleet. Dahua WizSense is the fastest PPE path — gloves still need a field check.

VMS platforms

10

Analytics add-ons on existing VMS. Medium–high agnostic fit; PPE/smoking coverage varies.

Ipsotek for enterprise compliance. BriefCam when investigation matters. IntelliVision for OEM embed.

Edge appliances

19

Turnkey boxes and NL/VLM platforms. Best path when you need mixed RTSP without an ML team.

Lumeo first for smoking+PPE POC. Coram/Spot for PPE/SOP. Dragonfruit / FuweeVision as smoking alternates.

Eliminated

7

Cloud-only, EOL, or wrong workload. Phase 5 re-checked all seven — none reinstated.

See the appendix. Full records kept; nothing dropped silently.

Comparison highlights

Pick YOLO11 for new projects; RF-DETR if Apache is required. Edge appliances are the highest commercial .

BackboneCOCO mAPLicenseVerify
YOLO1139.5AGPLpartial
YOLOv837.3AGPLverified
RF-DETR48.4Apachepartial
D-FINE54.0Apachepartial
RT-DETRv246.5Apachepartial
ApplianceNL/VLMNote
LumeoNoSmoking + gloves pages
Coram PointYesNative PPE Jul 2025
Spot AI IVRPartialPPE/SOP; smoking undocumented
Ambient PulsarYesNL events; unverified smoking
Scylla AsteriaNoSmoke ≠ cigarette

Score: 5 = best in catalog · 3 = viable · 1 = poor fit. ML ops: 5 = low burden.

OptionCostAccuracyAgnosticLatencyML ops
YOLO11 + DeepStream55551
RF-DETR + MMDeploy55542
Lumeo24534
Coram Point23545
Spot AI IVR23545
RK3588 + YOLOv8n53432
Dahua WizSense43155

Smoking & gloves

Smoking

Still true: no verified mAP on hold-out CCTV. Phase 4 conflated unverified with non-existent. Lumeo, Dragonfruit, FuweeVision, Visionify, and XINHUO document smoking/vaping modules. Scylla smoke ≠ cigarette. Coram HALO environment smoke ≠ cigarette. DIY: YOLO11n on UiTM (1940) + fire-exit arxiv (8124).

POC: Lumeo → Dragonfruit → FuweeVision/Visionify. DIY: YOLO11n on UiTM + fire-exit. NL (Coram / Ambient / Spot Iris) is experimental.

Gloves

DarthRegicid recall 0.555 is verified for that checkpoint only. itsadityabaniya Gloves mAP50 ~0.69 is unverified. SH17 includes Gloves with released weights; per-class metric TBD. Hexmon adds an explicit NO-Gloves class. Target ≥0.70 recall on your site. Commercial PPE often omits a glove-specific SLA.

Bench three OSS baselines plus Hexmon NO-Gloves. Do not treat 0.555 as a ceiling.

ViolationPretrainedTrainingData need
Helmet / vestYesRecommended fine-tune200+ site images
GlovesPartial (0.555 recall)Required for SLA500+ labeled
SmokingNo verifiedRequired500–2000 (UiTM = 1940)
Fire / smokeYes (mAP50 0.872)Optional100+ if scene mismatch

Hardware tiers

Workload: 640×640 detector, 1080p ingest, 15–30 FPS per stream. Do not compare batch numbers to e2e without normalization. Published matrix: 30 devices.

DeviceYOLOv8nYOLOv8sYOLOv8mFit
Orin Nano 8GB724858 (4 cam)≤4 cameras · kitchen pilot
Orin NX 16GB256196360 (12 cam)4–12 cameras · factory default
RK3588 SBC5328~12Budget ARM; m is not real-time
Coral Edge TPU60 @ 320px——Retrofit · accuracy trade-off
Hailo-8——8.15 e2eRPi5 add-on; hw_only batch ≠ e2e
Intel NUC OpenVINO304——x86 VMS analytics
Hi3519 / CV181x28 / 22— / 7—On-camera OEM · budget IPC
1–2 cam
YOLOv8s / YOLO11s

Orin Nano, RK3588, Coral

3–4 cam
YOLOv8s / YOLO11m

Orin Nano (tight) or Orin NX

5–12 cam
YOLOv8m

Orin NX 16GB

12+ cam
YOLOv8m + batch or x86 GPU

dGPU NVR, NUC, Hailo multi-stream

Full device grid →

Agnostic architecture

Any IP camera → edge box. Swap models without rewriting ingest.

Ingest

Any IP camera over RTSP/ONVIF. Decode with GStreamer or DeepStream.

MAL

Boot-time backend: TensorRT, RKNN, HailoRT, OpenVINO, NCNN. Versioned artifacts per class.

Violation engine

Per-class thresholds (gloves lower than helmet). Temporal confirm on smoking.

Alert / VMS

Normalize events. Hybrid: Coram/Spot webhooks while custom models train.

Counts are live. Phase 4 reported partial 61; the catalog now has more partial rows after audit additions.

  • Verified 3 — Source cross-check matched. Safe to cite in procurement.
  • Partial 70 — Cited from vendor or README; not re-run internally.
  • Vendor claim 3 — Marketing metric. Do not treat as mAP.
  • Unverified 29 — No public data. Require a vendor POC.
  • — Mean average precision — standard object-detection accuracy on a labeled set.
  • — Copyleft license on Ultralytics YOLO. Closed SaaS typically needs an enterprise license.
  • — Real-Time Streaming Protocol — the usual IP-camera video feed.
  • — Contracted performance target, e.g. recall on your cameras.

Risks and open items

POC itemAcceptance
Internal e2e RTSP benchSame 1080p workload; p50/p95 latency
Smoking POC (Lumeo / Dragonfruit)Written detection rates on your cameras
Triple glove baselineDarthRegicid · itsadityabaniya · SH17 — recall ≥ 0.70
Vendor SLADemo on your glove and smoking scenes
License review~15 HF models marked “Not stated”
NL vs deterministicLumeo pipelines vs Coram/Ambient NL
RiskSeverity
Glove false negatives at distanceHigh
Smoking false positives (steam, shadows)High
AGPL on YOLO11/v8 in closed SaaSMedium
Vendor metric inflationMedium
Coral/Hailo accuracy vs GPUMedium
Ambarella CV72S — no public YOLO FPSMedium

Phase 5 audit

9
Missed solutions added
2
Phantom / stale URLs
0
Wrong eliminations
12
Shortlist challenges

Added: Staqu JARVIS · FuweeVision · Visionify · XINHUO · TensorVok · Cogniac · SH17 YOLO weights · Smoking DAHD paper · Fire-exit CCTV set.

Phantoms: TaroPlay/Smoking GitHub 404. HF imnuman/fire-detection-yolo 401 — use GitHub as canonical.

Upgrades: Lumeo #1, Dragonfruit, SH17, FuweeVision. Downgrades: Coram/Spot for smoking, Ambient NL as default, DarthRegicid as sole glove floor.

Eliminated (7)

All seven re-verified. No reinstatements. Actuate “edge” is a camera filter plus VPN to cloud.