Progress this week
This week we got all three workflows running individually with live camera input on the Jetson. Each can now capture and process incoming frames rather than rely on saved video. We have not yet run all three together or completed the live camera-to-tablet path.
Harry filmed longer rack videos for offline testing before moving to the live feed, then began profiling memory usage, latency, and throughput. Alan got the treadmill workflow running live and started checking processing speed and update delay, including the effects of sampling and voting. Jerry moved the bench workflow to live input, continued testing camera alignment and plate association, and worked on connecting observations to the shared timers and display.
Risks and responses
Running all three workflows together is still our biggest technical risk. The individual runs give us a starting point, but we need to see how processing time and memory use change under the combined workload. If it is too slow, we will first reduce unnecessary model calls or stagger station checks. Lower resolution and optimized models are other options, but we will check their effect on accuracy before adopting them.
We also need to make sure live processing does not fall behind the camera. Next week we will measure observation age through the tablet path and check that stale or disconnected inputs show unknown.
Bench accuracy and full five-minute unattended events still need more testing. We also need footage kept separate for final evaluation. UC recording permission remains pending, so we are continuing to use the apartment gym.
Design changes
There were no major design changes this week. Live camera input was part of the existing plan. We will use the combined Jetson measurements to decide whether we need changes to model settings or how often stations are checked.
Schedule
Individual live inference is now running, but combined operation and the tablet connection are still behind our earlier plan.
- Week of November 2: run all three live workflows together, measure combined performance, and connect their outputs to the tablet.
- Before the November 9/11 interim demo: measure camera-to-tablet delay, test camera interruptions, and rehearse the demo.
- After the interim demo: fix remaining integration problems and complete longer unattended-event and held-out accuracy tests.