Skip to content
Independent Research Project · AI Computer Vision for Workplace Safety

Seeing floor hazards before anyone falls.

FloorEye is a self-funded research and engineering project by Junaid A. Shah. It applies computer vision to cameras a business already owns, identifies spills and wet floors as they appear, and triggers a warning automatically. The goal is simple and public in nature: fewer slip-and-fall injuries in the workplaces that can least afford them.

Sole inventor and developer U.S. provisional patent filed Running on live camera feeds
Filed
U.S. Provisional Patent
819+
Documented R&D Hours
193
Pages of Documentation
TEST STORE CAM-01 LIVE
FloorEye
Camera frame from a test store showing FloorEye identifying a water spill on the floor
WATER SPILL IDENTIFIED
4 / 4 STAGES
Warning Light On Staff Notified Event Logged
Continuous frame-by-frame analysis
Works on existing IP and CCTV cameras
Mapped to OSHA and NIST requirements
Built and funded entirely independently
The Problem

A spill nobody sees is an injury waiting to happen.

In a busy store, kitchen or warehouse a wet floor can appear in seconds: a leaking cooler, tracked-in rain, a dropped drink. It sits there unnoticed, and somewhere inside that window a worker or a customer goes down. Today most businesses find out only after the fall.

People get hurt

Falls on the same level are among the most common and most preventable injuries suffered by American workers and customers.

Detection depends on luck

A hazard is found when a person happens to walk past it. That is minutes of exposure on a floor where seconds matter.

The cost lands on small employers

One fall can mean a workers' compensation claim, litigation and higher premiums for years, a burden a small business absorbs far less easily than a national chain.

Compliance is hard to evidence

OSHA expects floors to be kept clean, dry and hazard free. Proving that a business actually did so, minute by minute, is nearly impossible with paper checklists.

Existing technology skips the small operator

Advanced safety monitoring has largely been built for enterprise budgets and new hardware, leaving independent businesses with nothing but a mop and a cone.

Nobody can watch every aisle

A manager cannot monitor every floor, every aisle, every minute of a shift. The exposure never actually goes away.

240,000+
Slip and fall injuries requiring days away from work each year in the United States
$54,499
Average workers' compensation cost per slip or fall injury claim
$165,514
Maximum OSHA penalty per willful or repeated violation
#2
Same-level falls rank second among the costliest causes of disabling workplace injury

Why this problem is worth solving at a national scale

Same-level falls are a multi-billion dollar annual cost to the U.S. economy and a leading source of lost workdays. The businesses most exposed, independent retailers, restaurants, small warehouses and service operators, are precisely the ones that cannot fund enterprise safety systems. FloorEye is built on the premise that meaningful injury prevention should not require new cameras, a capital project or a corporate safety department. It should run on the equipment a small business already has.

How It Works

Turning cameras that record the past into cameras that prevent the next injury.

FloorEye connects to existing camera streams and adds a layer of machine perception that does not get distracted, does not take breaks and does not walk past a puddle. When a hazard appears, the system identifies it, warns people nearby and records the event automatically.

Stage 01

Observe

Frames are sampled continuously from each camera at the edge and screened for spills, standing water, freshly mopped surfaces and caution signage.

Stage 02

Verify

A candidate hazard passes through a multi-stage verification pipeline, including comparison against a learned dry-floor reference for that exact camera view, before anything is raised. This pipeline is the subject of the provisional patent filing.

Stage 03

Warn

Connected warning lights and signage switch on through IoT integration and staff are notified by app, SMS and email. No human has to notice first and no human has to press anything.

Stage 04

Record

Every hazard, alert and resolution is timestamped and stored, producing an objective record of how quickly the condition was found and cleared.

Reactive cleanup becomes documented, automatic prevention.

Engineering summary

Designed, built and deployed end to end by a single developer.

Edge
Python inference agents running on compact edge hardware inside the store, with local RTSP ingestion and ONNX model execution. Video is processed on site rather than shipped wholesale to the cloud.
Platform
FastAPI backend, MongoDB, Redis and Celery, MinIO and S3 object storage for event clips, encrypted tunnel connectivity between edge and cloud.
Interfaces
React and TypeScript web dashboard, React Native mobile application, on-demand live video streaming, push and SMS notification delivery.
Perception
Instance segmentation model trained on floor imagery collected in real environments, with per-class confidence thresholds rather than a single global threshold, plus a self-learning dataset pipeline that accumulates labeled examples toward a fully proprietary model.
Privacy
The system is designed so that stored and displayed evidence is anonymized. Hazard detection does not depend on identifying any individual, and no personally identifiable imagery is retained.
Real Detections

Not a concept. Running against live camera feeds.

The frames below were captured from live cameras in participating test stores with the FloorEye model running. No staging and no stock photography. Real cameras, real floors, hazards identified automatically.

TEST STORE CAM · CHECKOUT
FLOOREYE
HAZARD Camera frame from a test store showing a water spill identified near a checkout counter

Water spill identified at checkout

A fresh water spill on the floor near the register was segmented and flagged from the camera feed, without any person reporting it.

Water SpillLive FeedAutomatic
TEST STORE CAM · AISLE
FLOOREYE
MONITORING Camera frame from a test store showing caution signage and a recently mopped floor area being tracked

Caution signage and mopped floor tracked

The model distinguishes caution signage from the wet surface itself, which is what lets the system tell a marked hazard apart from an unmarked one.

Caution SignMopped FloorMulti-class

Captured under signed written authorization from the participating store owners covering camera access and data use.

Field Testing

Validated in real stores, not only on a benchmark.

Two independent retail locations are currently running the system on their own cameras under signed authorization, deliberately configured to test the same claim from two directions.

Test Site A

Three live camera streams

Running on a compact ARM-based AI edge module installed on site.

Existing store cameras, no new camera hardware
Continuous operation across normal trading hours
Detection events feeding the training dataset
Test Site B

Three live camera streams, different hardware

Running on a standard low-power x86 edge computer rather than an AI accelerator.

Chosen specifically to prove hardware neutrality
Same software stack, same model, different silicon
Removes dependence on any one vendor's hardware
Why two different machines matters. A safety system that only works on one expensive piece of hardware cannot reach the businesses that need it most. Running the identical stack on two deliberately different edge platforms demonstrates that the approach travels, which is the difference between a laboratory result and something that can be adopted broadly across small and medium-sized businesses. Both deployments are test installations of this personal research project, operated with the written consent of the store owners.
Research, Patent & Documentation

The record behind the work.

FloorEye was conceived, researched, engineered, deployed and documented by one person. The evidence of that work is itself part of the project.

Intellectual Property

U.S. Provisional Patent Application

Multi-Stage Verification System and Method for Detecting Wet-Floor and Liquid-Spill Hazards from a Fixed Camera Feed.

Application No. 64/073,277
Filed May 23, 2026
Sole inventor and applicant: Junaid A. Shah
Official USPTO filing receipt issued July 10, 2026
Independent Investment

Self-funded research and development

Every hour and every dollar of development, from first prototype to production deployment, was funded personally with no outside capital, grant or employer support.

Development period May 2025 to July 2026
Documented effort 819+ hours
Funding source Personal
Technical Record

193 pages of original documentation

A complete written technical record covering system architecture, technology stack, installation methodology and end-user operation, authored alongside the build.

System Architecture Reference
Technology Stack Manual
Administration Manual
New Site Setup Guide and Owner Manual
White Paper
Seeing Hazards Before People Fall
Junaid A. Shah · Independent Research · May 2026

Seeing Hazards Before People Fall

An AI Computer Vision Approach to Real-Time Wet Floor and Spill Detection for Small and Medium-Sized Businesses

A published technical paper examining how modern computer vision can identify spills and wet floors the moment they appear, why small and medium-sized businesses carry a disproportionate share of slip-and-fall risk, what federal regulation actually requires of them, and how an automated perception layer closes that gap. The paper maps the system directly to OSHA Walking-Working Surface standards and to NIST SP 800-53 security and privacy controls.

AuthorJunaid A. Shah
Pages11
PublishedMay 2026
FocusSafety & Compliance
Regulatory Alignment

Safety you can actually evidence.

The OSHA Walking-Working Surfaces standard requires employers to keep floors clean, dry and free of hazards, and to be able to show it. Paper logs cannot demonstrate what a floor looked like at 2:14 in the afternoon. A continuous, timestamped detection record can. FloorEye was designed from the outset around that evidentiary gap.

Automatic audit trail

Every hazard, alert and response is timestamped and stored, available for an inspection or a claim without anyone having to reconstruct it later.

Documented speed of response

The record shows when a hazard appeared, when it was flagged and when it was cleared, which is exactly the standard of care regulators look for.

Objective evidence in disputes

An impartial, contemporaneous record of monitoring is stronger than recollection for every party involved in a slip-and-fall dispute.

Mapped Against

Published safety and security standards

OSHA Walking-Working Surfaces (29 CFR 1910 Subpart D)
NIST SP 800-53 Security and Privacy Controls
Encrypted data handling, footage remains the site owner's
Continuous, timestamped event logging

The full mapping of the system to OSHA and NIST requirements is set out in the published white paper above.

About

One researcher, one system, built end to end.

FloorEye is not a company and has no team. It is a personal research and engineering project.

Junaid A. Shah JS

Junaid A. Shah

Sole Inventor, Researcher & Developer

Technology professional with more than seven years of experience across data engineering, business intelligence, machine learning and full-stack software development, holding an M.S. in Business Analytics. FloorEye began as an independent question about whether the cameras already hanging in ordinary businesses could be made to prevent injuries rather than merely record them, and became a working, deployed system.

Every element of the project is his own work: the research and problem framing, the perception model and its training data, the multi-stage verification method now under provisional patent, the edge software, the backend platform, the web and mobile interfaces, the field installations and the full written technical record.

Sole inventor on U.S. Provisional Patent Application 64/073,277
Sole author of the published technical white paper and the 193-page documentation set
Sole developer of the edge agents, backend platform, dashboard and mobile application
Self-funded throughout, with no external investment or institutional backing
LinkedIn
FAQ

Questions, answered.

What people most often ask about the project.

No. FloorEye is an independent research and engineering project developed and funded personally by Junaid A. Shah. There is no corporate entity behind it, no investor and no team. The deployments described on this site are test installations operated with the written consent of the participating site owners.
No. The system reads existing IP and CCTV cameras that support standard RTSP streaming, which covers the large majority of installed systems. Avoiding a hardware purchase is central to the point of the project, since the businesses most exposed to slip-and-fall risk are the least able to fund one.
Detecting water on a floor from a fixed camera is difficult because wet surfaces are largely defined by subtle changes in reflection rather than by shape. The contribution of this work is the multi-stage verification method, which compares a candidate hazard against a learned dry-floor reference for that specific camera view and passes it through successive validation gates before raising an alert. That method is the subject of the provisional patent application.
The system monitors walking surfaces continuously and produces a timestamped record of each hazard and the response to it. That record evidences the documented, diligent monitoring the Walking-Working Surfaces standard contemplates, and it exists whether or not anyone thought to write in a log book that day.
Video is processed on site at the edge rather than streamed wholesale to the cloud, and connections are encrypted. Footage remains the property of the site owner. Only event evidence and metadata are retained, and stored or displayed frames are anonymized. Hazard detection does not depend on identifying any individual. The design is mapped to NIST SP 800-53 security and privacy controls.
Work in progress includes converting the provisional patent application to a non-provisional filing, replacing the current perception model with a fully proprietary one trained on the accumulating field dataset, and expanding field testing to additional sites so the approach can be validated across a wider range of floor surfaces, lighting conditions and camera placements.
Contact

Questions about the research?

For technical questions, collaboration on field testing, or requests for the full documentation set and patent record, get in touch directly.

Email Junaid A. Shah
Sole
Inventor & Developer
Filed
U.S. Provisional Patent
Published
Technical White Paper
Deployed
In Real Environments