2.6 KiB
Intermittent person lock — 2026-09-27
Joshua reported that the camera finds a person but does not stay locked. He confirmed he was in the camera view during the live trace. The camera's person detector alternated PERSON and NO_PERSON roughly every 250 ms, and many bounding boxes were anchored to the top-left image edge (for example, box=0,0,95,124 with score 0.433). The current three-frame gate clears immediately on NO_PERSON, which explains the TRACK/IDLE flicker. It also selects the highest-scoring box independently each frame, so it can jump between detections.
A temporary JPEG diagnostic build produced frame 41, frame 61, frame 81, and the paired log. The image shows mostly ceiling/window, with only part of Joshua at the far left edge; the view also appears sideways relative to an upright person. The detector's edge boxes match this poor framing. This is the primary issue to correct before tuning temporal tracking, because a hold filter would otherwise preserve an unreliable target.
The diagnostic JPEG build was removed, and the previous normal camera application (SHA-256 0acc763624b20a09afbf512e20afad51ffc9de4276102ffa639a0345d39d70a5) was rebuilt, flashed, and hash-verified. No target-selection code changed in this diagnostic pass.
Next: re-aim/rotate the camera so a person's head and upper body are upright and near image center at the intended approach distance. Capture a new image and 15–20 second movement trace. If detections still intermittently drop with good framing, add short miss hysteresis and spatial association to the existing target-selection logic, then retest empty-scene false positives and target departure.
Re-aimed retest
Joshua re-aimed/rotated the camera and stood in view while moving left to right. In a 20-second three-board serial capture, the camera reported PERSON on all 80 results and NO_PERSON on none; 77 commands were gaze_valid=1, with three brief gaze_valid=0 results from the current position-consistency gate. Both eye consoles received matching valid gaze sequences. Joshua then confirmed both eyes stayed on TRACK and followed throughout the movement. This resolves the observed bench lock loss by correcting the camera view, without changing the tracking code. Installed-position walk-bys and empty-scene false positives still need testing. A future hold/association filter should be judged against those tests, not inferred to be needed from the original poorly framed view alone.