Trial use of AI camera Task Cam for sequential assembly verification and step timing for automotive parts manufacturers in Thailand

How the brief came about

The challenge in this case is to know whether employees assemble parts through every step in sequence, and how many seconds each step takes. The customer is an automotive parts manufacturer in Thailand. Assembly at the station involves several steps. The customer wants a time chart for each step, and video clips to review afterwards.

In 2026, we installed Task Cam in Sequential mode alongside a second Task Cam camera in Skeleton mode on a live assembly line at this automotive parts manufacturer.

Requirements and constraints

The system must check the work sequence by detecting employees' hands under real site conditions. The main requirements and constraints are as follows.

• Employees wear white fabric gloves, and the existing hand detection model did not detect hands in this type of glove accurately.
• Employees work quickly, so the camera must detect hands in time at every step.
• Video clips and data files of every step must be stored on a NAS (network-attached storage device).
• The customer wants to use a single tablet on site. Showing images from both cameras side by side on this tablet is still an ongoing trial in the POC round.

How does Task Cam check that employees complete steps in sequence?

Task Cam checks the sequence from the employee's hands entering the detection frame (an area of the image that the camera is set to watch) of each step. These frames are drawn at the points where hands must reach in to pick up or assemble, arranged in the actual work sequence, and the camera then compares what the employee does with that sequence in every work cycle.

We chose to check by detecting hands because the challenge is the order in which parts are picked up and handled, and we use a second camera in Skeleton mode because the customer also wants the time of each work cycle. The Duo view web page shows images from both cameras side by side in a web browser.

1. 1

   Define steps on screen

   Draw the detection frame for each step in a web browser, arranged in the actual assembly sequence.

2. 2

   Show the next step with frame colours

   A yellow frame shows the step to do next. A green frame means the hand has entered the correct frame. A red frame is a step whose turn has not yet come.

3. 3

   Count pieces when the sequence is complete

   The system increases the piece count only when the employee has completed all steps in sequence.

4. 4

   Time with Skeleton mode

   A second Task Cam camera in Skeleton mode reads the body skeleton and times each work cycle.

5. 5

   Store clips and files for review

   The system records video clips and CSV files (table data files that can be opened in Excel) separately by step, then stores them on the NAS.

To solve the glove issue, we used real fabric gloves from the site to retrain the hand detection model, then applied this model to the cameras on the production line.

What was delivered in the POC round

In the POC round, we delivered a work sequence checking system at a live workstation, together with a data storage system and a user manual. The items are as follows.

• Task Cam in Sequential mode, which checks the sequence by detecting hands and counts pieces
• A second Task Cam camera in Skeleton mode, which times each work cycle
• A hand detection model trained for white fabric gloves
• Storage of video clips and CSV files of every step on the NAS
• A Duo view web page that shows images from both cameras side by side in a web browser (side-by-side viewing on a single tablet is still an ongoing trial)
• User manual for the on-site team

What stage is this work at, and what comes next?

This work is ongoing as a trial on a live production line in the POC round. The next step is a production version of the software with cameras of a higher processing specification. The sequence of work in 2026 is as follows.

1. Meeting with the customer to set the approach. The conclusions included a hardware upgrade and retraining the model for white fabric gloves.
2. Project start, purchase order (PO) received, and system installed at the workstation
3. User manual delivered to the on-site team
4. Trial on the production line, with periodic progress reports to the customer
5. Next step: a production version of the software with cameras of a higher processing specification, and a code change so that the cameras can find the NAS on their own

Confirmed results during the trial

So far, the confirmed results are that the system has been installed and is being trialled on a live production line, and that we have delivered the user manual to the on-site team. There are no measured figures yet because the POC round is ongoing, so we report only what is confirmed by documents.

Problems found during the trial are documented. Employees' hands moved faster than the camera model used in the POC round could detect in time, which is the reason for using cameras of a higher specification in the production version. For the problems on site, we applied interim fixes by replacing the camera that had a problem and requiring the NAS to be ready before the cameras are switched on.

What we learned

The key principles from this work are to test the model with real gloves, to set the camera specification from real hand speed, and to define the NAS power-on sequence from the start. Details of each are as follows.

• Test the model with what employees actually wear. White fabric gloves meant the existing hand detection model did not detect accurately, so we collected images of the gloves on site to retrain the model before use.
• Set the camera's processing specification from actual hand speed on site. The POC round is the period for measuring whether the camera can detect in time with the work pace, before selecting hardware for production use.
• A system that records to network storage must have its power-on sequence defined from the start. If the cameras start before the NAS is ready, recording will have problems.

How can this approach be applied to your assembly line?

If your assembly line has a fixed sequence of steps, and skipping a step causes defects, you can use the same approach as this case: draw detection frames according to the work sequence, check by detecting hands, and store clips and the time of each step. What needs adjusting is the hand detection model, the camera specification and the installation position, according to the conditions at your station.

• Data from Skeleton mode should be used to adjust workstations and work sequences, not to assess individual employees, and employees must be informed before the cameras are installed.
• Detection frames in the same image should not overlap.
• Start with the station that has the clearest work sequence problem. Record observations on hand speed, lighting and gloves, and only then set the specification for other stations.