Robovision and delaware automate safety with deep learning

Nov 08, 2019
  • operations
  • discrete manufacturing

Cloud might be the hot term of the decade, but as data processing technologies evolve, exciting niche applications of on-premise AI systems are also emerging. On that note, the team recently joined forces with AI platform developer Robovision with the goal of using deep learning to tackle a human-centered problem: ensuring the safety of workers in risky production environments without going to the cloud for data processing power.

  • Challenge: Developing a deep learning system that assists with locating workers in the manufacturing environment, to ensure workplace safety.
  • Solution: A smart algorithm that identifies people in video on a real-time, frame-by-frame basis.
  • Result: An extremely fast solution that very accurately tracks people, even if they are at the edge of a zone or walking next to each other.

“A customer came to us with a really interesting challenge,” begins Wouter Labeeuw, Manager Data & AI at delaware. “Their workers were using and working around heavy machinery in specific areas, and they wanted to accurately count the number of employees entering and exiting these zones, no matter the circumstances.”

Safety monitoring: “a task more suited to the right machine”

Workplace safety is the number one priority for manufacturing companies. “Knowing where people are in facilities is crucial – not just because people could be in danger, but also to ensure the safety of firefighters and other first responders in the case of an emergency,” Wouter explains.

Our customer was relying on a basic electronic system that counted every time a laser beam at foot level was interrupted by a person passing through it. Wouter: “There’s an obvious blind spot here – the system gets confused when two or more people walk through at the same time or pass in and out of the zone by wandering around its boundary. To solve that problem, they had somebody monitoring the zone via camera.”

Time-consuming, repetitive, manual tasks like screen monitoring require lots of focus over a long period of time and can lead to fatigue and human error. “It’s definitely a task more suited to the right machine, which can efficiently support the human monitor,” Wouter asserts. “It was up to the delaware and Robovision team to develop and train that machine.”

Directing digital eyes to dangerous areas

The dual-company team crafted a smart algorithm that identifies people in video on a real-time, frame-by-frame basis. “It compares changes in frames by measuring the distance between two points. In doing so, the algorithm can accurately identify people in images and track their movements.”

But first, the system had to be trained to recognize people. “In this facility, employees working in high-risk areas wear very specific clothing,” Wouter goes on to say. “So, the algorithm orients its tracking to that clothing. In the future, we’d like to expand its repertoire with additional training data.”

People bunching up together? Workers lurking around the edge of a zone? People who almost cross the boundary but reenter? The system accurately tracks them all – in real time.
Wouter Labeeuw, Manager Data & AI at delaware

Fully wired, self-contained and in real time

To make things more complex, the company wanted the system to be fully wired and self-contained. Wouter: “Working with machine-learning algorithms and high-speed cameras isn’t exactly light on data processing. To solve this issue, Robovision has developed an application that simplifies AI module training, shrinking the hardware footprint needed for real-time data processing. After all, real-time, high-frame-rate tracking is absolutely essential in a safety context – every millisecond counts.”

While the system is still a proof of concept, the Robovision-delaware team has successfully created a solution that is extremely fast and highly accurate. “People bunching up together? Workers lurking around the edge of a zone? People who almost cross the boundary but then go back in? The system doesn’t lose track.”

Even more impressive is the fact that the system processes images from no fewer than 12 cameras. “There are lots of different ways that employees can enter or move around monitoring zones. The end goal is to create a system that sees and tracks with such coverage and accuracy that imperfect human monitoring is no longer a factor. Our module constantly drinks in image data and learns from it through the input of a human monitoring expert.”

learn more

In search of an intuitive, efficient way to get computer vision up and running in your production environment? Learn more about our Intelligent Vision Kickstarter, powered by Robovision. Or would you like to automate and optimize your company’s quality inspection process? Go to the delaware store and discover Quality Control, powered by Robovision

related content