Computer Vision Solution for a Large Industrial Client
Employee safety is an important focus of every industrial plant which makes it essential for companies to track the compliance of personal safety. Many organizations are putting forth a significant effort into implementing automation systems that use AI to track and document whether employees are complying with personal safety requirements. This is a complex effort which involves many studies of AI. First Line Software has already taken part in one of those studies.
The Task
The client requested that First Line use the existing live video streaming capabilities of its CCTV cameras as the foundation for implementing an AI capability to monitor whether the company’s safety requirements were being met. This required that the AI system be able to identify specific plant production lines, employees, and key safety elements.
The Process
The First Line Software engineers created a convolutional neural network and taught it to identify the specific staff members, safety equipment elements (helmets, safety jackets, etc.) and the production line types, using the live stream from the CCTV cameras.
The pilot version of the system can track and react to the three most common scenarios of human behavior:
- Whether the staff is wearing the necessary safety equipment (helmet and safety jackets)
- Whether the individual has covered his helmet with a hood (which is strictly forbidden)
- Whether an individual has attached a rope when performing high-altitude work
Dataset
The typical problem of this type of research task is the lack of exact cases and the scenarios which can be used to “teach” the neural network. In order to solve this challenge, First Line had to create all the potential positive and negative scenarios of an individual’s behavior. The process also involved analyzing and creating 56 models of human behavior that could be displayed on-site at the plant.
As you can see in the photos, some of the employees were fully equipped and met the safety requirements, while some part of employees did not. The system uses 12 pivot points to analyze each individual and additional pivot points to analyze the safety equipment. Each video frame has a specific color identification as well as the text frame. In addition, the viewer can see the in-plant location trigger.
Object recognition
The algorithm developed by First Line analyzes the data in three steps. The first step is to analyze the CCTV captured frame and analyze the human presence. If a human is detected the algorithm pushes the frame to the convolutional neural network. Once this step is completed the network uses the pivot points to identify the person and the safety equipment elements – helmets and/or a safety rope. Using the “support vector” algorithm, the program compares the captured image of the person with the scheme, which is stored in database. If the algorithm detects that the safety rules have been broken, it transmits a message to the management staff.
The technology
In order to achieve the best results, First Line used Mask R-CNN (Detection Platform) to analyze the captured frames. This framework was chosen because it was the best fit for the required task and can identify and highlight objects with the frames. We taught the neural network using the transfer learning script. This script was chosen since there are no requirements to collect statistics about the completed work or the number of workers at the plant, and no need to identify the specific department where each employee spends the majority of their time at work.
The result
First Line was able to achieve the sustainable analytics of the live video with the ability to detect the objects and classification of the same. The detection ratio varies from 77 to 100 percent. The pilot project showed great results and the customer is currently running in-debt testing on their premises.
Download the case study by clicking the PDF icon on the left-hand side above.
Download Case Study: Computer Vision Solution
Download Case Study
related projects

Visual recognition mobile app for counting forestry materials
The goal of a timber holding company was to improve accounting and control systems for lumber production by precisely counting their in-stock building materials.
Show details
Image Recognition Solution for a Media Company
Our customer asked our engineers to develop an online system that could identify and recognize military equipment in a photograph.
Show details
Machine Learning for Data Governance in a Hospital Setting
Wide-spread adoption of EHR systems brings along large amount of data available for reporting and other analytical insights. Health organizations routinely generate thousands of such analytical artifacts on daily or weekly basis.
Show detailsContact Us
USA
Cambridge MA
1 Broadway,
14th Floor,
Cambridge MA 02142
USA
San Mateo CA
400 Concar Dr,
San Mateo California
94403
Netherlands
The Hague
Louis Couperusplein 2,
4th floor 2514HP,
The Hague
Australia
Sydney | Brisbane
12 Creek Street,
Brisbane QLD 4000
United Kingdom
Gloucestershire
Cowley House,
12 Black Jack Street Cirencester
Gloucestershire, GL7 2AA
Czech Republic
Praha
Na Hřebenech
II 1718/8,
140 00 Praha 4
Czech Republic
Brno
Centrum, Šumavská,
Šumavská 416/15,
602 00 Brno

Send us a note
We'll do our best to answer within one hour