Privacy-friendly People Counting using Machine Learning
The goal of this project was to use machine learning to develop a self-contained (no cloud) people counting system using only low-cost and privacy friendly sensors (i.e. no cameras, no radar, etc.) that runs on a low power, ceiling mounted sensor cluster.
I came up with the proof-of-concept, pitched it to management, and then led the development and successful deployment of this new type of occupancy estimation system.
The technology eventually led to a patent application. You can read more about the project in this whitepaper that I authored, which gives an overview of the technology and it's applications.
This was a huge project that I really enjoyed working on. There's something about bringing new technology, all the way from ideation to production that is very fulfilling to me. This project had many challenges but in the end the feature delivered as it was an industry-first privacy-friendly people counting solution at this price point.
My Role #
As the senior ML engineer and technical lead for this project, I was responsible for the following:
- Model development, training pipeline, and versioning
- Distributed data collection system and data pipeline
- Automatic data labelling pipeline using computer vision and transfer learning
- Productionizing the model and inference pipeline including realtime data acquisition, quantization, and optimization for deploying on a resource constrained embedded system (IoT device).
- Monitoring and continuous feedback system
- Coordinating with other teams and people within the organization as well as external parties
- Leading and collaborating with the software development team to implement the inference pipeline on the target hardware
- Coordination, documentation, and ownership transfer with the product team
- Developed and released a new, low cost, privacy preserving, self-contained, approximate people counting solution that's accurate enough many practical applications
- The feature increased the demand for the O3-Edge product
- Resulted in new intellectual property