Accomplished
This week I moved the treadmill workflow from recorded footage to a live camera feed on the Jetson. The board now processes incoming camera frames and updates treadmill states while the camera is running.
– Live input: I used recorded-video runs first to check the deployment, then switched to the camera. This let me examine capture and processing together, including whether the workflow keeps up with incoming frames.
– Performance: I began profiling memory usage, per-check processing time, and throughput. I also looked at the parts of the update delay separately. Capture, sampling, inference, and voting all contribute, so timing the pose model alone would not tell us how quickly a live change is detected.
– Difficult views: I continued checking cases where someone is partly hidden by the console or standing beside the belt. These cases also matter when evaluating lower resolution or frame skipping. I have not adopted either as the final configuration because we still need to compare their accuracy and delay.
Schedule status
Live treadmill inference is now running on the Jetson. The remaining integration work is getting those live states onto the tablet and checking performance while the rack and bench workflows run on the same board.
Next week I’ll prioritize the combined run and measure the time from a live step-on or step-off event to the corresponding tablet update. That will give us a measurement of the complete path before the interim demo.
Deliverables for next week
– Run all three live workflows together on the Jetson with Harry and Jerry.
– Measure step-on and step-off delay through the tablet display.
– Compare processing time and observation freshness when running alone and under combined load.
– Check blocked views and camera interruptions.
– Evaluate lower-resolution or frame-skipping changes against the current configuration.