Tracking Node Catalog
Tracking category
Generated from 3 catalog nodes in AI/ML/Tracking.
Nodes in this category
Showing 3 of 3 matching nodes.
Associate Entities
AI/ML/TrackingGives local tracks from one or more cameras a global identity by comparing appearance embeddings. A track is collected for a few observations before it creates a new entity; later tracks that look like a known entity are matched to it, even on another camera. Every board and user of this app using the same task id shares the identities. They live in this process's memory only: after an hour without calls or a restart the task starts over and entity ids begin again at 1. An app can have at most 16 active tasks (64 per process); an idle task is released after an hour.
Extract Appearance
AI/ML/TrackingCrops every detection from the frame and runs a re-identification model on the crops in batches. The built-in models are Intel OpenVINO person re-identification models (Apache-2.0): the chosen one is downloaded into Cache Dir on first use, checked against its SHA-256 and read from there afterwards. With Model set to custom, connect your own session from Load ONNX and set Normalization to match its export: imagenet for torchreid/OSNet exports, raw for FastReID onnx_export.py exports, which normalize inside the model. Each detection becomes an observation with an L2-normalized appearance embedding, keeping its box, camera, session, tracker id, track id and state (track-only fields such as hits and velocity are not carried). Lost tracks from Track Detections (Include Lost) carry a predicted box, so they are passed through with their existing embedding instead of being cropped. Feed the result into Track Detections for appearance-aware tracking or into Associate Entities to recognize the same object across cameras. Boxes that are not finite or smaller than Min Box Size are left out.
Track Detections
AI/ML/TrackingFollows detected objects across the frames of one camera and gives each a stable track id (ByteTrack, with BoT-SORT appearance matching when detections carry embeddings). Run it once per frame: the tracker is kept in memory per user, board, node, Camera pin and Session pin between runs and uses frame timestamps for motion and expiry. Track ids are only unique per tracker instance, named by each track's tracker_id: a new Camera or Session pin value, 10 minutes without frames, eviction when too many trackers are open, or a process restart starts a new tracker with a new tracker_id (a session carried by the detections does not). Deployments that spread runs over several processes keep one independent tracker per process.