Tracking one person across cameras
Tracking one person across cameras is the reconstruction, after the event, of where a single subject went on a site, from the camera that first saw them to the one that saw them leave. Camzify does it from a description rather than a face. Attribute extraction records what each tracked person wore and carried, AI suspect search finds every appearance that matches, and the cross-camera journey map orders those appearances into one timeline with the coverage gaps shown.
- Search by description, not by face
- Every appearance across every camera
- One timeline with the gaps shown
Camera by camera, an afternoon per person
A report comes in after the fact: a person in a gray jacket took something from the back of the store at around two. The investigator has a time, a camera and a description, and everything else is manual. They scrub the first camera to find the person, guess which neighboring camera they walked into, pull that feed, scrub again, and repeat for as long as the path continues.
On a site with forty cameras that is the slowest part of the investigation, and it is easy to get wrong. A missed hand-off is a gap in the account, and a wrong guess is an hour on the wrong feed. The evidence often sits in the recordings and is never assembled, because assembling it costs more time than the case is thought to be worth.
Law enforcement has the same problem at larger scale. Digital forensic examiners told a National Institute of Justice workshop that they deal with as much video evidence, particularly from surveillance cameras, as they do with phones, with limited tools to analyze it (RAND Corporation for NIJ, 2015).
- 01CAM 01Main gate2 checks
- 02CAM 04Loading dock2 checks
- 03CAM 09Server room2 checks
- 04CAM 02Parking lot A1 checks
A description becomes a search, and a search becomes a route
Every tracked person is described as they are seen. AI attribute extraction attaches structured attributes to each detection, clothing color and type, carried objects and a plain-language behavior note, so the footage is indexed by what was in it rather than only by camera and time.
- AI suspect search takes a plain-language description, gray jacket and dark backpack, and returns every matching appearance across the indexed cameras and the time window, ranked by confidence.
- The cross-camera journey map is built from one of those matches and orders every camera that saw the same subject into a single timeline, with each hand-off timestamped.
- Where the subject left camera coverage, the timeline shows the last confirmed appearance and the gap, rather than guessing a path.
- Each hop carries a confidence score, and the investigator confirms or rejects it, so the route on the page is one a person has checked.
- The route exports with its frames for an incident file or a handoff to police, and the recordings behind it stay in video backup for the retention period.
None of this is face recognition. The match is made on clothing, carried objects and timing, the same details a witness would give, and the identification of a person is a human decision. The feature pages say so and it applies here.
AI attribute extraction
Clothing, carried objects and a behavior note attached to every detection, which is what makes the footage searchable by description.
DetectionAI suspect search
A plain-language description returns every matching appearance across the cameras, ranked by confidence, with no photo required.
DetectionCross-camera journey map
The matched appearances ordered into one timeline, hand-off by hand-off, with the coverage gaps shown rather than guessed.
DetectionMulti-object tracking
The confirmed tracks that the description, the search and the map all rest on, kept apart through crowds and crossings.
- Entrance camera unobstructedCompliant
- Sales floor camera in focusCompliant
- Stockroom door in frameNot compliant
- Rear exit camera litCompliant
A failed item is resolved as Fixed or Pending before the round can close.
The round that keeps the evidence usable
A journey map can only pass through cameras that were recording and pointing the right way, so the patrol round on a site that expects to investigate is a camera-readiness round. At each stop the checklist asks whether the view is unobstructed, in focus and lit, and whether the door or aisle the camera was installed for is still in frame, because a display moved in front of a lens is otherwise found in a search three weeks later.
A failed item messages the person designated for that camera, and the report carries the frame. Camera health monitoring covers that round in detail, and incident investigation covers what happens once the frames are pulled.
A check found Not Compliant captures a snapshot and messages the guard designated for that camera; on a manual round the operator chooses to send it, on an automated round it goes on its own. The item stays Pending until it is marked Fixed, which captures the after frame, and the report shows both.
Why this matters, in published figures
Figures from the retail industry body, the Congressional Research Service, the FBI and a peer-reviewed study of camera evidence. Camzify publishes no figures of its own.
- $112.1 billion
- Retail shrink in 2022, up from $93.9 billion in 2021, in the National Retail Federation's 2023 National Retail Security Survey of 177 retail brands. Internal and external theft accounted for 65% of it.
- Source: NRF, 2023 National Retail Security Survey
- 36%
- The share of overall inventory shrink that the 2023 survey's respondents attributed to external theft including organized retail crime, as summarized by the Congressional Research Service in May 2024. The same report notes that the FBI's crime data collects shoplifting but does not capture ORC specifically.
- Source: Congressional Research Service, R48061
- 15.9%
- The share of reported property crimes cleared by arrest or exceptional means in 2024, against an estimated 5,986,400 property crime offenses, in the FBI's Uniform Crime Reporting summary released August 2025.
- Source: FBI UCR, Reported Crimes in the Nation, 2024
- 45% and 29%
- In a peer-reviewed analysis of 251,195 crimes recorded on the British railway network between 2011 and 2015, CCTV was available to investigators in 45% of cases and judged useful in 29%, and useful CCTV was associated with significantly increased chances of the crime being solved (Ashby, 2017).
- Source: European Journal on Criminal Policy and Research
What it will not do
A search across cameras is a tool for a person building a case, and its limits are worth knowing before the case depends on it.
- It will not recognize a face or confirm who a person is; it matches clothing, carried objects and timing, and a change of jacket between cameras breaks the match.
- It will not tell two people in the same uniform apart with certainty, and the confidence score on each hop says so.
- It will not follow a person through an area with no camera; the timeline shows the gap and does not guess.
- It will not search footage that was never recorded or that has passed its retention window.
- It will not decide anything; every hop is confirmed or rejected by the investigator, and the export is what they confirmed.
We do not publish match rates or search times. The trust page sets out why.
Industries where this applies
Frequently asked questions
Ready to patrol your site 24/7?
Book a 15-minute demo and see a live patrol run on your own cameras.