Computer Vision
AI Car Wash Monitoring: How Cameras Track Every Vehicle and Service
How AI car wash monitoring uses your existing cameras to log check-ins, service times and cleaning steps, and what operators gain from the data.
- Published
- Updated
- 7 min read
Table of Contents
- Why is it hard for car washes to measure their operations?
- What does AI car wash monitoring do?
- What does one session record look like?
- How does AI car wash monitoring work?
- What did we learn building it?
- How does AI compare with manual logs and POS data?
- What can operators do with the data?
- Do you need new cameras?
- How do you get started?
- Get a free car wash assessment
Running a car wash looks simple from the outside: cars go in dirty and come out clean. Inside, the business depends on things that are hard to see without data: how many vehicles really come through, how long each one takes, which bay or shift is slowing things down, and whether every service step was actually done. AI car wash monitoring uses the cameras most sites already have to answer those questions automatically.
This article explains how it works, using the system we built for a car wash site as the example. You can see that system's real output in our car wash analytics case study.
Why is it hard for car washes to measure their operations?
Most operators rely on manual logs, POS totals or a manager's gut feel. Manual logs are incomplete on busy days. POS data shows what was sold, not how long it took or whether it was done properly. And without a reliable picture of busy hours and service times, staffing and pricing decisions are guesswork.
There's also a trust problem. When staff report the counts, owners have no way to check them, so under-reporting goes unnoticed. AI gives an independent count of every vehicle, straight from the camera.
What does AI car wash monitoring do?
A computer vision system watches the camera feed and turns it into structured data: which vehicle arrived, when it checked in and out, how long it took, and which service stages were seen.
- Detects and classifies each vehicle as it arrives.
- Records check-in and check-out for every car.
- Measures service time per vehicle, per bay and per shift.
- Marks each service stage the site defines (such as water, soaping, washing, external and internal wash) only when the camera sees it.
- Produces a processed video with detections drawn on it, so any vehicle can be reviewed later.
The result is a dashboard that shows the day car by car, instead of a pile of footage nobody has time to watch.
What does one session record look like?
Each car becomes one session: vehicle type, check-in, check-out, total time and the stages that were seen. In one recorded session from our project, an SUV checked in at 20:56:22 and out at 21:05:12, a total of 8 min 50 sec, with 3 of 4 service stages detected. In another, a sedan (classified with 0.95 confidence) took 8 min 02 sec with all 4 stages seen.
Put a day of those records together and you have what manual logs never give you: how many cars came through each bay, how long each one really took, and which paid stages were skipped.
How does AI car wash monitoring work?
The system processes video from live camera streams or uploaded recordings. A detection model finds vehicles in each frame, a tracker follows each vehicle through the site so it keeps one identity from entry to exit, and event logic turns those movements into check-ins, check-outs and service stages. A stage that isn't seen on camera is reported as not detected rather than assumed. Results are saved to secure cloud storage and shown on a web dashboard, with an API for other systems. The case study shows each of these outputs on a real wash bay.
The hard part isn't detecting a car; it's doing it reliably on real footage: different camera angles, wet lenses, night lighting, vehicles partly hidden by equipment or staff. That's why every site starts with a short assessment of its own camera views, and why the training data comes from the site's own cameras, spray and mist included.
Under the hood, video comes in from RTSP camera streams or uploaded MP4 files. A YOLO detection model (PyTorch) finds and classifies each vehicle, ByteTrack keeps one ID per car from entry to exit, and a custom-trained model plus zone logic recognises service and cleaning steps. Processing runs on NVIDIA GPUs with OpenCV and FFmpeg, a FastAPI backend exposes the results as a REST API, and footage is kept in encrypted AWS storage behind role-based access. See the full stack on our car wash analytics page.
What did we learn building it?
Three design decisions made the data trustworthy enough for an owner to act on. Each came from a real constraint on site.
- Tick a stage only when it's seen. A report that fills in expected steps looks complete but can't be trusted, which is the same problem manual logs have. In one recorded session, internal wash wasn't seen, so the panel shows 3 of 4, not 4 of 4.
- Make every number traceable to footage. A quality check or a customer complaint needs more than a number in a report, so the system writes a processed copy of the video with its detections drawn on it, next to the session record.
- Classify the vehicle, don't identify it. Bays are filmed with customers and their cars in view. The system records the vehicle type (SUV, sedan) and doesn't read number plates; plates are pixelated in our demo video, and footage access can be limited to authorised staff.
How does AI compare with manual logs and POS data?
Manual logs and POS data tell you roughly how many cars came in and what was sold. AI monitoring adds what actually happened: when each car arrived and left, how long it took, and which stages were done. It's also the only one of the three that doesn't depend on someone writing it down.
| Manual logs | POS data | AI camera monitoring | |
|---|---|---|---|
| Vehicle count | Often incomplete on busy days | Paid transactions only | Every vehicle the camera sees |
| Service time per car | Rarely recorded | Not recorded | Check-in to check-out, per car |
| Service steps done | Staff say so | What was sold, not what was done | Each stage marked only when seen |
| Can the owner verify it? | Hard to check | Matches the till, not the bay | Reviewable video for every car |
| Extra work for staff | Entries every shift | None | None |
What can operators do with the data?
The data turns daily decisions that used to rely on estimates into ones based on counts and times you can check. These are the five most common uses.
- Plan staff around real busy hours instead of estimates.
- Find bottlenecks: the bay, shift or service that takes longest.
- Check quality: confirm every paid service step was completed.
- Compare sites in a multi-location business on the same numbers.
- Price services based on how long they actually take.
Do you need new cameras?
Usually not. Most systems work with existing cameras as long as the view covers the entry, exit and bays clearly. A quick review of a sample video shows whether the current setup is enough, or whether one camera needs moving.
How do you get started?
The simplest way to start is with recorded video from one bay. It shows what the AI can measure on your site before connecting live cameras or adding more locations.
- Send a sample video from one bay, a few minutes long, from the camera you already have.
- We assess the view: whether entry, exit and the service stages are visible, and in what lighting.
- Pilot on recorded video from that bay, so you see real session records before anything is installed.
- Connect live cameras once the results match what happened on site.
- Add bays and locations, with every site measured the same way.
Get a free car wash assessment
Want to see what AI picks up at your car wash? See the car wash analytics case study, then send us a short video from one bay for a free assessment.
Hikmat Ullah
Computer Vision Engineer, ML Bench
Hikmat Ullah is part of ML Bench's computer vision team. He works on video analytics for real camera footage, including the car wash monitoring system described on this blog.
Technically reviewed by Dr. Waqas Sultani, CTO
