Counting Towels, Capturing Time: Video AI for Manufacturing Line Visibility

Quick Summary

Challenge
A towel manufacturer needed to track throughput and idle time at stitching stations, but had no way to make sense of their camera feeds.
Solution
Tatras Data built a deep learning pipeline using object detection, pose estimation, and action recognition to count towel folds and monitor worker activity in real time.
Result
92% of workstation activity captured and integrated into SAP

Tech Stack

AI: Pose Estimation Action Recognition Object Detection | ML: Custom neural networks for temporal action segmentation | Data & Retrieval: Real-time video from stitching tables | Dev: Python PyTorch OpenCV Custom heuristics pipeline | Infra: On-premise video analytics integrated with SAP | Security: Local-only video processing with role-based access

The Challenge

The client (a high-volume textile manufacturer) had invested in camera systems across their factory floor. But those feeds sat underutilized.

They couldn’t see how many towels were being folded per station. Nor were breaks monitored. Supervisor walkthroughs were infrequent and subjective.

Management wanted real metrics:

  • How many units per hour?
  • Who’s idle, and when?
  • Which stations consistently underperform?

Manual tracking was impossible at scale. They needed eyes that didn’t blink, and brains behind the lens.

A Day in the Life: Before Our Solution

Every shift ended with a number — total towels folded.

But how did that number come to be? Somehow, no one really knew.

If one station underperformed, managers were left guessing:

  • Maybe the folder took an unscheduled break
  • Maybe the towels were bulkier
  • Maybe the camera just didn’t catch the right angle

Supervisors walked the floor when they could, but it was impossible to track every table, every hour.

The video feeds were there; but reviewing them was tedious, expensive, and too slow to act on.

There was no way to measure individual throughput. No way to catch inefficiencies in real time. And, no way to know who was working, who was waiting, or why production dipped.

Decisions were based purely on gut feel. And performance blind spots only grew with scale.

Pain Points:

  • No automated way to count throughput per table
  • No system to detect idle time or break duration
  • Limited visibility into operator activity and role segmentation
  • Underused video infrastructure
  • Manual tracking too slow and inconsistent

Solution

1. Core Innovation

Tatras Data built a computer vision system tailored to the stitching workflow:
  1. Detected and localized each workstation using object recognition
  2. Used pose estimation to identify towel folders and their actions
  3. Classified sequences of movement as towel folds or idle periods
  4. Trained a custom neural network to recognize folding even with partial occlusions
  5. Synced output directly to SAP for centralized planning and reporting

2. Key Features

  • Video-based towel count tracking per workstation
  • Real-time action recognition of folding motions
  • Employee break detection via pose inactivity
  • Supervisor and loader role segmentation
  • Integration with ERP systems for live reporting

3. Workflow Integration

The video feed now does more than record.

Each workstation is monitored in real time. Not by humans, but by intelligent models.

Throughput per station, worker activity, and break durations are all quantified and sent directly to SAP.

Supervisors get a data-backed view of performance. While, ops teams get the clarity they need to optimize shifts.

Outcomes

✅ 92% of operator activity accurately captured 🧮 Towel count and break time auto-logged by workstation 🎯 Greater accountability and shift optimization 🔄 Seamless SAP integration with minimal manual input 👁️ Reliable visibility into true line productivity

Ready to build your AI system?

Let's discuss how our pipeline can accelerate your path to production.

Start a Conversation
You're interacting with a beta version of our chatbot—thanks for helping us improve!