How Azure Percept is Democratizing IoT
Review and Use Case for Azure Percept Developer Kit
I unboxed the Azure Percept Developer Kit and was able to quickly build a low code image recognition model that recognized if I was wearing a hard hat (Classification was performed on the edge instead of in the cloud) and then sent the result to Azure (IoT Hubs)
It’s pretty cool stuff, but before I jump into the tech, I’ll walk through why platforms like these are essential for digital transformation.
Outline
- Motivation: Why make a simplified end to end IoT platform?
- Usage: How does Azure Percept work?
- Example: Hard hat Detection Project
Motivation: Why make a simplified end to end IoT platform?
To make digital transformation (Especially in Manufacturing — Industry 4.0) occur, you need to enable both your developers and your business stakeholders. Streamlined end to end IoT platforms enable developers to present demo-able solutions to internal customers and the market. They also enable business stakeholders to demystify much of IoT and understand the art of the possible.
Enabling developers to create proof of concepts faster allows the rest of the company to react and provide feedback faster as well. This build momentum for iterative design. The ultimate solution may not leverage the simplified end to end tech from the initial solution, but it serves as a method to quickly create a testable demonstration to drive go-no go criteria for further development. Too often, businesses can get stuck reducing latency, refining environments, and optimizing interdependencies only to later realize there was no underlying business case for the proof of concept.
Enabling business stakeholders to work in a simplified sandbox demystifies much of IoT (And broadly, Industry 4.0). It’s one thing to send over a short deck of IoT and AI talking points and design constraints, but it’s something else entirely when the business makes realizations such as the following:
- “It seems like better training data improves the model more than additional model complexity”
- “How can we simplify the tool as much as possible initially to drive better adoption?”
- “How can we get our IoT data and systems (Batch) data all in one place to make better analyses with all our information?”
Realizations like these are pretty straightforward, but when businesses start to drive this kind of thinking, it removes the wrong assumptions of previous digital transformation thinking that a few IoT devices will instantly improve overall enterprise performance. Providing an end to end platform to do this with relatively little code enables these realizations to occur.
How does all this business and tech context connect Azure Percept? Developers spend less time thinking about infrastructure to build the initial proof of concept, and business stakeholders can finally get their hands dirty in building a proof of concept to both learn the tech and experiment with use cases to refine requirements to the developer team.
Usage: How does Azure Percept work?
Microsoft’s Azure Percept is a move to provide an end-to-end IoT solution for advanced functionality including image detection, audio recognition, and more. It accelerates the development path for edge-enabled AI use cases.
- The Overview: https://azure.microsoft.com/en-us/services/azure-percept/#product-overview
- The Docs: https://docs.microsoft.com/en-us/azure/azure-percept/
Once you’ve gone through the initial installation and setup, you can start building your proof of concept on Azure. A quick rundown before we use any more buzzwords:
- IoT (Internet of Things): Physical objects (Sensors and more) connected to the Internet
- Edge: The majority of the analytics are done on the device — not the cloud. The results are sent to the cloud but not the raw image and audio information. This both speeds up compute time as well as reduces cloud costs
- AI (Artificial Intelligence): Training a model to make decisions. Typically, people use the general term (AI) when they’re actually referring to specific field like Machine Learning (Primarily focused on classification, regression, and clustering algorithms)
Now let’s combine all that: Azure Percept enables developers to quickly deploy machine learning models (Templated or custom) to an IoT device on the edge. This enables a model to arrive at results (Such as image classification) faster and use less compute time and storage cost on the cloud.
Example: Example Hardhat Detection Project
For an initial project, I worked with some friends to demonstrate a basic image classification model. We picked a simple use case in manufacturing safety: Detecting if people are wearing hard hats on the shop floor. Shop floors can be dangerous! Bureau of Labor Statistics calculated ~3 incidences per 100 workers for the manufacturing sector.
Developing the model on Azure Percept involved the following:
- Start a new Azure Percept project from Azure Portal
- Capture training images (Stream images from the camera)
- Label images (Hard hat or no hard hat)
- Test the model on Azure Percept
- Deploy the model on the edge
- Stream the results (e.g. from IoT hub to a dashboard or storage device via Azure Stream Analytics)
Below is a high-level overview of the architecture used:
The results can easily be sent to a storage account (e.g. Azure Data Lake or Blob Storage) as well as streamed directly to a dashboard (e.g. PowerBI); this presents a wide range of opportunities. Think about the use cases: A shift supervisor can see historical data of the team’s adherence to safety policies. A supervisor can also be alerted immediately if an employee is not wearing a hard hat in specific areas as well.
Below are some of the outputs of this rapid proof of concept. This includes both streaming results from Azure Percept as well as a rough cut dashboard
Future Outlook: Enabling an enterprise to build low code proof of concepts such as this provide the opportunity for developers to move faster as well as for business stakeholders to demystify some of this technology. If you can get stakeholders excited about wearing hard hats, there’s many more use cases to show the art of the possible.
