Skip to main content

Client Meeting - 24/10

· 2 min read
Dan Robb
Team Member

Notes from our client meeting

Date: 24-10-2025
Time: 9:00 - 10:00
Team: Alina Zubova, Dan Robb, Henry Edwards, Talal Aljallal, Tom Lam, Michal Berkasiuk

Key Project Learnings & Context​

  1. Filter Gap (Last Year): A critical lesson from the previous project was the absence of a final filter on the results. This suggests a major area for us to improve upon in our pipeline.
  2. AppImage Binaries: These are essentially the compiled code for the radar chip. Their job is to convert the raw radar information into the specific data format we need for processing (e.g., point cloud).
  3. Chirp Configuration: This refers to the specific kind of chirps (radar signals) the hardware sends out, which directly affects measurements like range. Good news: it's built into the chip, so we shouldn't need to worry about configuring it ourselves.

Minimal Viable Product (MVP) Focus​

The MVP goal for this term is two-fold:

  1. Replication + Improvement: We need to reproduce the previous team's output but rely only on the radar unit for the data source.
  2. ML Demo: Get a basic demonstration working that involves training a model. This means we need to immediately focus on dataset handling and collection.

ML System Requirements​

  • The trained model should be running locally on our machines.
  • The model must be built using the PyTorch framework.

Data Collection & Visualisation Strategy​

Data Types​

The raw data we collect will be a mix of:

  • Point Cloud Data: A set of 3D points representing detected objects.
  • Tracks Data: Time-series data tracking the movement of detected objects.

Visualiser Goal​

If we develop a visualiser, a high-value feature would be to superimpose the radar information (e.g., detected points/boxes) directly onto the live video feed for immediate context and verification.

Recording Specifics​

  1. Radar Height: Keep the radar unit fairly low during recording, roughly 1m off the ground.
  2. Movement: When collecting data, ensure people walk right-to-left AND left-to-right in front of the radar.
  3. Final Classification Filter: We can use data like height and speed as a crucial final filter to help classify or distinguish different objects/people.