Project Skybird.
Built on Kundi · Deployed in government
Person detection, tracking, and trajectory analysis for video. Project Skybird identifies a target across frames, recovers from occlusions when subjects pass behind obstacles, identifies what caused each occlusion, and emits structured trajectory data for downstream analysis — designed to run inside government and security environments.
Persistent identity across frames, even through occlusion.
Skybird's ensemble combines a state-of-the-art person tracker, an open-vocabulary scene detector, and an occlusion-aware track-merge layer — producing browser-ready annotated video plus per-frame structured output.
-
Person tracking with ReID
YOLO11 + BoT-SORT with native re-identification for consistent tracking of a target person across frames, including across short disappearances.
-
Multi-person annotation
All detected persons get labeled bounding boxes with unique IDs in the output video — not just the target.
-
Occlusion recovery
Merges broken track fragments when a person passes behind an obstacle (tree, pole, vehicle, structure), preserving identity continuity through gaps.
-
Obstacle identification
YOLO-World open-vocabulary detection identifies what caused each occlusion ("tree trunk", "parked vehicle", "wall"), enabling explainable track recoveries.
-
Open-vocabulary scene detection
Periodic rescans catch scene objects entering the frame: vehicles, structures, street furniture, foreground items, traffic infrastructure.
-
Proximity alerts
Configurable spatial alerts — e.g., the screen flashes when the target enters a defined proximity to a detected scene object.
-
Trajectory output (JSON)
Per-frame bounding boxes, centers, timestamps, and confidence scores emitted as structured JSON for downstream analytics, dashboards, or replay.
-
Browser-ready video
H.264 + faststart encoding for direct browser playback of annotated output without external transcode steps.
See it inside your environment.
Skybird walkthroughs are available under NDA. Schedule a deployment assessment to see the live tracker, occlusion-recovery output, and structured trajectory data running against representative video sources.