Summarize Tracked PPE

How it fits

Typical backends

Caveats

• Hard constraintThe three input lists are consumed positionally and must have equal length; a frame with mismatched lengths stops the worker with an error, and with it the deployment, rather than being truncated or answered empty.
• Hard constraintEach classifier probability vector must hold exactly two entries, the missing-PPE one at missing_class_index and present-PPE at the other; any other length stops the worker with an error.
• AccuracyA track with no usable crop yet scores exactly prior, so a prior inside the uncertain band keeps an unseen person uncertain instead of present.
• AccuracyA track's running probability is weighted by body-part detection confidence and the unfilled history slots are padded with prior, so a person just entering view leans toward prior until enough confident crops accumulate.
• I/O contractCrops below the configured minimum width or height are dropped before reaching a track's history, so a person seen only at small scale can sit in the uncertain band because no sample qualified.
• Parameter interactionThe status has three bands (present / uncertain / missing); collapsing or inverting positive_confidence_threshold against negative_confidence_threshold removes the uncertain band and lets the missing-PPE verdict take precedence.
• State lifetimeA track's history is dropped once its last usable crop is more than forget_num_frames frames old, even while the person is still tracked at a size below the minimum; keep forget_num_frames at least as large as the upstream tracker's own forget window.
• State lifetimeConfiguration is read once at startup, so a change takes a redeploy. An empty frame emits an empty list and still counts toward forget_num_frames.