Accomplished
I started this week by going through the design report feedback that applies to the treadmill workflow. I made the treadmill behavior more specific: when a reading counts as unknown, what starts and stops the unattended timer, and how we separate measured laptop results from planned Jetson behavior. Some feedback asked for clearer justification of using a pose model instead of simpler image methods. To answer it, I wrote up the comparison I ran before the report and explained how the delay in the pose model’s updates affects how fresh the treadmill status is.
I also began moving the treadmill workflow onto the Jetson Orin Nano. I set up the environment and dependencies for the pose model and started running it on recorded treadmill footage. The aim is to confirm that the board gives the same results as the laptop before testing with a live camera. Alongside this, I looked at ways to reduce the update delay, including skipping frames and lowering the input resolution, so we have options ready if the treadmill check turns out too slow on the Jetson.
I also worked with Jerry on hard treadmill views, such as a person partly blocked by the console or standing next to the treadmill instead of on it. In those cases the workflow should return an unknown reading rather than a confident wrong answer.
Schedule status
Setting up the Jetson is progress on the hardware work we planned for the week of October 19. However, I have not finished running the treadmill workflow on the board, and live-camera testing and integration are still behind our earlier schedule. I also have no performance measurements on the Jetson yet.
To catch up, next week I will first get the treadmill workflow running on recorded footage on the Jetson and measure it, then move to a live camera. Harry and Jerry are doing the same for their workflows, so all three of us will have individual baselines before we run everything together.
Deliverables for next week
- Run the treadmill workflow on the Jetson with recorded footage and confirm it matches the laptop results.
- Measure memory usage, per-check latency, and throughput, and check whether the update delay still meets our freshness target.
- Run the treadmill workflow with a live camera feed on the Jetson.
- Keep testing difficult treadmill views and confirm that unclear cases return unknown.
- Work with Harry and Jerry to run all three workflows together on one Jetson and measure their combined performance.