The challenge
Our client builds car wash tunnels used across the US. Inside a tunnel, cars ride a conveyor. If a driver brakes or shifts gear, a car can roll back into the one behind it. Operators had no automatic way to see this, and by the time someone noticed, the damage was done.
The goal
Detect a rollback or a likely collision as it starts, and stop the conveyor automatically. Record everything so each event can be reviewed later.
What we built
- A custom trained YOLO model detects and tracks every vehicle on the conveyor from the tunnel cameras.
- Speed, distance and direction are measured for each car, frame by frame, to tell a normal stop from a rollback.
- When risk crosses a threshold, the app talks to the tunnel PLC over Modbus and halts the belt.
- A desktop operator app shows live video, alerts by priority, event history and PDF reports.
- Video is recorded by session, time, count or schedule, and events are uploaded to cloud storage with email alerts.
The hard parts
Every site has different cameras, light and tunnel layouts.
Per site calibration and a config sync so each tunnel can be tuned remotely.
The system must run inside the tunnel, not in the cloud.
FP16 models on NVIDIA Jetson Orin NX, shipped as an ARM64 package with a one step device setup.
False stops cost money and annoy customers.
Trajectory checks over several frames before any stop command is sent.
The result
The system has been in production since August 2024 and has gone through many releases as new sites came online. It ships as a Windows installer for operator PCs and as an edge package for Jetson devices.