M.Video-Eldorado: in-store video analytics — an honest look at a pilot, not a 'chain-wide rollout'

• Видеоаналитика М.Видео: архитектура (Raspberry + RTSP-фильтр + YOLO), сценарии, пилот в 1 магазине, алерты 25→5/день — Habr (блог М.Видео-Эльдорадо), 2021-03
• Видеоаналитика М.Видео-Эльдорадо: 30 000 камер, один компьютер и нейросеть (хронология июль–октябрь 2020, команда 2–4 человека) — Retail.ru, n/a
• М.Видео-Эльдорадо масштабирует систему видеоаналитики (50 магазинов, 250 камер, конверсия +35% к сопоставимым, уведомления −75%) — М.Видео-Эльдорадо (пресс-релиз), 2021-08-06
• «М.Видео-Эльдорадо» масштабирует систему видеоаналитики (точные формулировки: «рос на треть быстрее», 5 локаций в июне 2021) — Retail.ru, 2021-08-06
• Группа М.Видео-Эльдорадо: итоги 2024 (более 1200 точек, +100 магазинов, 25 новых городов) — CNews, 2025-01-16
• Информационные технологии в М.Видео-Эльдорадо (ИТ-бюджет 2026 — 9 млрд ₽) — TAdviser, n/a
• Группа М.Видео-Эльдорадо объявляет о назначении CIO (Александр Соколовский) — ComNews, 2020-08-03

Background

M.Video-Eldorado is Russia's largest consumer electronics retailer: over 1,200 stores as of early 2025; in 2024 the chain added 100 stores and entered 25 new cities, with GMV up 5% (CNews). The technology base matches the scale: the group's 2026 IT budget is 9 billion rubles (TAdviser), and the IT function has been led since August 2020 by CIO Alexander Sokolovsky.

The company began building its video analytics system in the summer of 2020 with its own data office — no off-the-shelf platforms. The decision was both principled and forced: the chain has more than a thousand stores with about 30 surveillance cameras each, putting the target scale at roughly 30,000 cameras. Replacing the entire camera fleet to fit a vendor solution's requirements is economically unrealistic at that count; the system had to be built on top of the heterogeneous 'zoo' of CCTV cameras already hanging from the ceilings. The starting position, recorded in the company's Habr blog, was notably honest: nobody could estimate the economic effect in advance — there were only optimization hypotheses, and the only way to test them was a pilot.

The industry context sharpens the story. Retail video analytics is one of the most overheated segments of the AI market: vendors promise 'double-digit conversion growth' out of the box, while most publicly visible deployments never get past a few-store pilot. Against that backdrop, M.Video-Eldorado chose the opposite strategy — build rather than buy, with a tiny data-office team of two to four people, but on an architecture designed from day one for 30,000 cameras.

What makes this case interesting is precisely its documentation at every stage: a detailed engineering breakdown of the pilot on Habr and Retail.ru (March 2021), then a scaling press release with the first business figures (August 2021). Across these two genres — engineering and corporate — you can watch the same project sound different, which is itself a useful lesson in reading AI case studies.

Problem

An offline store is 'blind' compared to an online one: nobody knows where shoppers go, where they get stuck, how long they queue, and how many leave without ever reaching a consultant. Sales staff physically cannot see every visitor, and classic door counters cannot answer what happened to a person between the entrance and the checkout. For an electronics store this is especially painful: appliances are a consultation-driven product, and a visitor who has stood unattended at the washing-machine shelf for ten minutes is, with high probability, a lost sale.

The engineering side of the problem is no easier than the business side. Store cameras are heterogeneous, old, and installed for security rather than analytics; the standard configuration leaves blind zones, and some cameras had to be reoriented while new ones were added. The video stream from 30 cameras at 1080p and 25 fps is about 6 Mbit/s per camera: pushing that traffic from a thousand stores to a central cloud is impossible both bandwidth-wise and financially. Finally, the neural network must tell shoppers from staff, and camera images must be mapped to the store planogram — doing that manually takes far too long.

A separate fork was build versus buy. Ready-made video analytics platforms existed, but at 30,000 cameras the solution's economics are defined by cost per camera, and the company bet on in-house development by a tiny team — in the blog's own words, 'essentially students' working 'in hooligan mode'.

Solution

The architecture turned out emphatically cheap — and interesting precisely for that.

On the store side, each camera connects to an inexpensive 'Raspberry-family' device (about $50) that takes the RTSP stream and runs it through a home-grown filter: only keyframes are extracted — one frame every two seconds instead of 25 frames per second. As a result, traffic from a store's 30 cameras fits into roughly 3 Mbit/s — less than the raw stream of a single camera. Frames go to a private cloud in the company's own data center, where a single analytics engine based on the YOLO network (chosen for low resource consumption and retrained for the task) detects people, converts their positions into Cartesian coordinates, and maps them to the store planogram (Habr, Retail.ru).

The first-wave scenarios: the 'lonely shopper' alert — if a visitor stands or moves around the floor alone for too long, a message goes to the store's chatbot and a consultant goes to help; 'checkout queue' — a notification when shoppers in the pickup area exceed the norm; and the 'store heat map' — a density distribution of visitors for analyzing routes and merchandising. Distinguishing staff from shoppers was solved with almost comic pragmatism at first: people in red T-shirts (M.Video's uniform) are ignored; the 'elegant' solution via multi-camera tracking was deferred to the next stage.

The pilot started in one store in October 2020 — just three months after development began. A month and a half went into tuning the alerts: at first the system fired 25 'lonely shopper' alerts a day, most of them false — consultants would quickly have learned to skip such messages. By the end of the period, 5 alerts a day remained, which the company's blog interprets as 'a sharp rise in staff attentiveness': signals became rare and precise, making it meaningful to respond. In parallel the team solved earthly problems: blind zones, camera reorientation, manual planogram mapping — for scaling they wrote standard instructions letting store staff deploy the solution themselves in one day, plus a web portal automating the mapping.

In June 2021 the solution was set up in five more Moscow locations, and on August 6, 2021 the company announced scaling to 50 stores in Moscow and the region by year-end — with cameras connected to the neural network growing from 50 to 250. The press release called the project 'Russia's first video analytics system with proven economic efficiency' and valued the in-house build at one-fifth the cost of market analogues, with installations paying back within the first month.

Result

This case's results exist in two genres, and it is important not to mix them.

Genre one — the engineering blog (Habr, March 2021). The only impact metric published there concerns the alert itself: over 1.5 months of the pilot, 'lonely shopper' alerts fell from 25 to 5 per day; the system learned to filter false signals and bother staff only when it mattered. On business metrics the article is explicit: multi-camera tracking would only make them measurable — visit-to-purchase conversion, impact on average receipt.

Genre two — the corporate press release (August 2021). The figures there are bolder: 'the customer attraction coefficient in pilot stores grew a third faster', 'the conversion coefficient grew 35% versus comparable stores', notifications about helping a customer or opening checkouts fell 75%, installations pay back within the first month, and the in-house build costs one-fifth of market analogues. The release calls the project 'Russia's first video analytics system with proven economic efficiency'. The methodology of these comparisons (how many stores, what period, how 'comparable' ones were selected) is not disclosed, and there is no independent verification.

In our view, the scissors between the two genres are the most instructive part of the case. In March the engineers write 'we will only now learn to measure conversion'; in August the press office reports '+35% conversion'; strictly speaking, both texts can be true (pilot stores may indeed have outgrown a control group), but trust in a number depends directly on the genre it was published in. We would build a business case on the engineering version and treat the press-release one as an upper bound.

The second observation: even by the boldest public data, the project reached 50 stores out of 1,200+ — roughly 4% of the chain — and no public confirmation of a full-network rollout has appeared since. The distance between 'proven economic efficiency' in a pilot and total deployment is years and billions (30,000 cameras versus 250 connected at press-release time). That is why we deliberately keep this case at its real scale: a solid, well-documented pilot with early scaling — and a useful ruler for everyone reading other companies' 'chain-wide rollouts'.

Lessons learned

1. A pilot is not a rollout: 'works in 1 store', 'scaling to 50', and 'works in 1,200' are separated by different years and budgets; an honest case names its stage.
2. The real work in a CV project is alert precision: cutting from 25 to 5 alerts per day is the deployment, because staff stop ignoring the system.
3. Scale economics are decided by edge engineering: a keyframe filter on a $50 device squeezed 30 cameras' traffic into 3 Mbit/s — without it, a 30,000-camera project fails on bandwidth and money alike.
4. Build on existing infrastructure: the solution runs on the heterogeneous security cameras already hanging in stores — replacing the fleet would have killed the project before launch.
5. Pragmatic hacks beat perfect solutions at the start: 'ignore people in red T-shirts' is working staff identification for a pilot; multi-camera tracking can wait.
6. Distinguish reporting genres: the engineering blog publishes alert quality and says 'we will only now learn to measure conversion'; the press release says '+35% conversion' with no methodology; build your business case on the former and treat the latter as an upper bound.
7. Absence of rollout data is data too: since August 2021 no public confirmation of a chain-wide deployment has appeared, and the case must end where the sources end.

Frequently asked questions