Real-time Hard Hat Detection with Edge Computing and Deep Learning

As a fun demonstration project, we decided to train a hardhat detection model and run it on our self-cleaning camera. Because our cameras perform AI inference at the edge (ie directly on the camera as opposed to in the cloud or on a separate computer), the latency is very low. If you watch the video below you can see how quickly the camera detects when I remove my hard hat. This model is running at around 22 FPS. We have the camera connected to a buzzer and you can hear the buzzer go off immediately when it detects when I am no longer wearing my hard hat.

There are several reasons why our cameras perform this fast. Because we are performing AI at the edge, there is very little delay and latency compared to performing AI in the cloud. This ultra fast response time is simply not possible if we were doing this in the cloud! Sending high-bandwidth camera data to the cloud introduces several different sources of latency from the encoding, decoding, and networking steps. Our cameras run 100% independently and can perform detections like this in the field without needing to connect to any external cloud or service. Even if you have poor network connectivity in your operations, our cameras will continue to perform with high accuracy and reliability.

In addition to edge computing, the Canopy Vision AI model training pipeline creates optimized lightweight models that do not sacrifice on accuracy. You can read more about our technology in our white paper!

This example is just one well-known use case of how this AI/Deep Learning technology can be used for business purposes. Some other really great examples of where Edge Computing and Deep Learning can have a huge impact are:

  • Detecting unsafe conditions and sounding an alarm to alert nearby workers
  • Detecting when a chute is plugging or overflowing and shut down the equipment feeding it
  • Detecting when an environmental or chemical leak is occurring to shut down the line and prevent further spills
  • Detecting equipment failures and triggering an action to prevent major downtime events
  • Analyzing product quality on a moving production line and triggering the necessary actions to remove off-spec product from the line without causing a shutdown

Want to get started on your own project? Contact us today! A project can be deployed in less than 10 days. You can see how it works here.

Leave a Reply