ZENITH: What needs action now

How it works

1. 01

   Align and check the records

   CIO corrects each device's timestamps for clock error and matches records to video, such as a cancelled sale to its clip, giving each match a probability. ZENITH flags likely empty shelves: items unsold for unusually long while records show stock.

   • Kalman filter
   • Probabilistic data association
   • Negative binomial model
2. 02

   Detect people, forklifts and smoke

   An on-site detection model finds people, forklifts and possible smoke. Pose locates the head and torso, where helmets and vests are checked over several looks; smoke is judged over a short clip, alongside your fire alarm, never instead of it.

   • Real-time detection transformer
   • Top-down pose estimation
   • Temporal shift module
3. 03

   Track on the floor plan

   ZENITH maps each camera's view to the floor plan in metres and follows people and forklifts anonymously, without facial recognition. Each person or forklift gets a probability of being inside a zone, allowing for the camera's margin of error.

   • Homography estimation
   • Kalman filter
   • Hungarian algorithm
4. 04

   Measure queues and machine states

   ZENITH turns customer arrivals, worked out from each checkout's receipts, and queue length into a waiting time. Where a machine's controller gives no running signal, the machine's electrical current is compared with running and stopped levels learned from its history.

   • Little's law
   • Gaussian mixture model
5. 05

   Combine evidence into events

   ZENITH combines probabilities of place, time and records into events, such as a restricted zone entered without a permit. Each camera's evidence threshold is tuned against false triggers like steam; a failed camera or late record never means all clear.

   • Product t-norm
   • Finite-state machine
   • Finite Markov chain imbedding
6. 06

   Choose which events become alerts

   ZENITH calibrates each event's chance of being real from alert reviews and a weighted sample of unsent events. Safety thresholds are tested against target miss and false-alarm rates, using staged events and random audits; a person decides what to do.

   • Platt scaling
   • Inverse probability weighting
   • Learn-then-Test

Published research

1. Events, not single images

   ZENITH builds an alert from an event on the floor plan rather than from a single image, for example a person who stays inside a restricted zone for several seconds. Reviews of video intrusion and smoke detection describe checks over time and position as a standard way to cut false alarms, and smoke detection runs alongside your fire alarm, never instead of it.

   • Lohani et al. 2022 Perimeter Intrusion Detection by Video Surveillance: A Survey. Sensors.
   • Çetin et al. 2013 Video fire detection – Review. Digital Signal Processing.

2. Each event with its record

   An alert is easier to act on when it comes with the record from the same moment: a cancelled sale with its video clip, an empty shelf with the stock count, a long queue with how many checkouts are open. ZENITH builds on research on joining data streams by the time each event happened and on studies that measured how queues and empty shelves affect sales.

   • Akidau et al. 2015 The Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data Processing. Proceedings of the VLDB Endowment.
   • Lu et al. 2013 Measuring the Effect of Queues on Customer Purchases. Management Science.
   • Chuang et al. 2016 On-Shelf Availability, Retail Performance, and External Audits: A Field Experiment. Production and Operations Management.

3. The simplest sensor that works

   A clamp-on sensor on a machine's power cable can show whether it is running, idle or stopped, with no camera needed. ZENITH builds on research into reading machine states from electrical current, which also finds that splitting one meter's reading between several machines is only moderately accurate, and more accurate the more machines are metered separately. It also builds on vibration analysis, which can show wear in a machine's bearings, and turns to cameras where records and simpler sensors leave gaps.

   • Hart 1992 Nonintrusive Appliance Load Monitoring. Proceedings of the IEEE.
   • Holmegaard & Kjærgaard 2016 NILM in an Industrial Setting: A Load Characterization and Algorithm Evaluation. IEEE SMARTCOMP.
   • Randall & Antoni 2011 Rolling element bearing diagnostics - A tutorial. Mechanical Systems and Signal Processing.

4. Fewer, trusted alerts

   In a controlled experiment, automation that raised many false alarms harmed people's performance more than automation that missed events, and hospital studies describe the same 'alarm fatigue'. ZENITH draws on alarm-management practice from the process industries, where each site sets how many alarms one person can handle in a given time, based on how long each alarm takes to deal with. In the same way, each team agrees how many routine alerts it can act on in an hour. Our own design adds two rules: safety alerts are never held back, and each alert can be marked done, not a problem or escalated, so a person decides.

   • Dixon et al. 2007 On the Independence of Compliance and Reliance: Are Automation False Alarms Worse Than Misses?. Human Factors.
   • Sendelbach & Funk 2013 Alarm Fatigue: A Patient Safety Concern. AACN Advanced Critical Care.
   • Dubois 2018 The Sense and Nonsense of Alarm System Performance KPIs: What Are Meaningful Values?. M.A.C. Solutions (UK) white paper.