Skip to content


PPE Detection

PPE Detection


PPE Detection

PPE Detection

Timelapse delivers a premium service offering time-lapse photographers from bespoke hardware to internal data management cleared for defense contracts. Their technology allows customers to clearly see what is happening on site, allowing them to monitor progress from their office or anywhere else in the world. This visual record of their customers' projects helps customers improve processes, monitor OH&S compliance and provides an invaluable record for dispute resolution.

Technology Stack


Timelapse produces a significant amount of photography data as part of their Cloud Render product. They wanted to offer their customers a way to help monitor their OH&S compliance without the cumbersome task of humans having to review the photography in entirety.  

Timelapse customers often need capability to monitor safety violations for: 

  • Employee Safety – ensure employees are operating safely 
  • Marketing – ensure any footage intended for public release does not contain violations 
  • Legal – ensure that sites are not breaching OSHA standards 


Blackbook was engaged by Timelapse to design, build and deploy a computer vision model capable of detection of OH&S breaches in their photography data.  

A Web Portal and custom AI object detection model was built on AWS cloud infrastructure, which together would assist with the safety review process. Note that a separate Web Portal with custom DNS was provided to each Timelapse customer as required. 

The AI computer vision model was trained to detect the absence of Hard Hats and Safety Vests, as well as the capability to detect social distancing violations (removed once COVID-19 restrictions were lifted). 

The Web Portal facilitates the submission of the Timelapse imagery for AI analysis. Once the imagery has been analysed, any resulting AI detections will be followed by a human-in-the-loop review. This stage allows for invalid detections to be deleted and any new detections to be created using an embedded annotation tool. Any corrections are captured for potential future stages of model retraining.  


The model achieved over 70% accuracy on the limited initial dataset with strong improvements with data diversification with Australian workwear data and model retraining through Web Portal use. The Web Portal and model are actively being used by Timelapse customers, with potential for product expansion.