Per-camera tracking that maintains consistent identities within each camera view, plus cross-camera appearance search (v1): pick a person and Visylix assembles their journey across your other cameras as operator-reviewed candidates.
Visylix Person Tracking fuses real-time object detection with Kalman-filter motion prediction and two-stage detection association (ByteTrack-style) to maintain consistent identities within each camera view as individuals move through a scene, reconstructing their movement paths through trajectory analysis. Beyond a single view, cross-camera appearance search ships today as a v1 capability: pick a person from any detection and Visylix finds appearance matches on the site's other cameras within a time window and assembles a chronological journey with travel times between cameras. This search is appearance-based (visual similarity) and returns candidates for operator review, not automatic identity assertions. It runs entirely on-premise, and a purpose-trained re-identification model for higher precision is in our training pipeline.
Core capabilities of the Person Tracking model.
Maintains a stable identity throughout a camera view, coasting through brief occlusions with Kalman motion prediction.
ByteTrack-style high/low-confidence association with Hungarian assignment keeps IDs stable in crowded scenes.
Reconstructs travel paths identifying entry/exit points, dwell times, and zone sequences for comprehensive movement analytics.
Automatically records transitions between custom-defined zones for heatmaps and flow visualizations across entire facilities.
Recovers tracks after brief obstructions using predictive motion models and configurable track buffers.
Tracks in real time on live streams and scales across many concurrent cameras per compute node.
Pick a person from any detection and Visylix finds appearance matches on your other cameras within a time window, assembling a chronological path with travel times between cameras. Appearance-based and operator-reviewed; a purpose-trained re-identification model is in development.
Pairs with Visylix semantic visual search to find people in recorded footage by natural-language description.
Real-world applications for Person Tracking.
Reconstruct where a person went across a site by reviewing appearance-match candidates on nearby cameras, with travel times, before acting on any identity.
Maps shopper journeys from entrance to checkout, measuring dwell times at displays and optimizing store layout for conversion rate improvement.
Tracks subjects of interest within each camera view with real-time alerts for restricted zone entry and loitering behavior.
Monitors foot traffic through choke points, corridors, and concourses to prevent dangerous congestion and optimize flow.
Analyzes occupancy patterns floor-by-floor for HVAC optimization, lighting automation, and space utilization planning.
Tracks passenger flow from check-in through boarding, identifying bottlenecks and optimizing gate assignments and staffing levels.
Performance and deployment details.
Add Person Tracking to your video pipeline in minutes.
Assign the model to specific cameras with zone definitions and sensitivity settings through the web UI or API.
The model processes video frames in real time, generating structured detection events with bounding boxes and metadata.
Receive instant alerts via webhooks, trigger automated workflows, or query detections through the REST API.
See how Person Tracking is applied across different sectors.
Explore other computer vision capabilities.
Talk to our team to see this model in action on your video feeds.
Within a single camera, Visylix Person Tracking fuses object detection with Kalman-filter motion prediction and two-stage (ByteTrack-style) detection association to keep a stable identity for each person as they move, coasting through brief occlusions and reconstructing their trajectory.
Yes, as a v1 cross-camera appearance search. Pick a person from any detection and Visylix finds appearance matches on your other cameras within a time window and assembles a chronological journey with travel times between cameras. It is appearance-based (visual similarity) and returns candidates for operator review, not automatic identity assertions. A purpose-trained re-identification model for higher precision is in our training pipeline. It runs entirely on-premise.
Yes. Per-camera tracking and cross-camera appearance journeys both run on your own hardware, with no data sent to any cloud, and work in air-gapped deployments.