AGENTS.md

AGENTS.md

Summary

This is the repository reference for person detection plus distance monitoring. Use it when you need person-to-person spacing logic on top of stereo-aligned detections.

Use This Example When

• You need social-distance or proximity monitoring.
• You want a bird’s-eye view derived from stereo depth.
• You need a person-specific spatial detection baseline rather than a generic detector.

Do Not Use This Example When

• You need general object spatial detections.
• You need hand-to-object safety logic instead of person-to-person distance.
• You need a pure RGB detector with no stereo branch.

Quick Facts

• Category: neural-networks/object-detection/social-distancing
• Shape: script+standalone
• Primary task: detect people and monitor their pairwise distance
• Entrypoint: main.py
• Standalone path: oakapp.toml
• Frontend: none
• Runs on: devices with CAM_A, CAM_B, and CAM_C; RVC2 peripheral, RVC4 peripheral, and RVC4 standalone packaging
• Requires: SCRFD person detector, stereo depth, and calibration
• Input: live color and stereo pair
• Output: Video, Detections, Distances, and Bird-eye view
• Models: SCRFD person-detection YAMLs in depthai_models/
• Visualizer / UI: DepthAI Visualizer via dai.RemoteConnection

Read First

• README.md
• main.py
• utils/measure_object_distance.py
• utils/host_social_distancing.py
• utils/host_bird_eye_view.py

Architecture

• A color camera on CAM_A feeds the person detector.
• StereoDepth on CAM_B/C aligns depth back to the RGB stream.
• DepthMerger attaches depth to the 2D person detections.
• Host nodes derive pairwise distances and a bird’s-eye view from those spatial detections.

Constraints

• The example requires three cameras and aligned stereo depth.
• It is person-specific; it is not a general spatial-object monitor.
• Distance logic quality depends on reliable depth for each detected person.

Related Examples

• neural-networks/object-detection/human-machine-safety: use this when you need palm-to-object safety logic
• neural-networks/object-detection/spatial-detections: use this when you need the general spatial-detection baseline
• neural-networks/counting/people-counter: use this when you only need a current-frame person count

Validation

• Run: python3 main.py
• Success looks like: the Visualizer shows people detections, distance annotations, and a bird’s-eye view that updates with scene geometry
• Common failure meaning: stereo alignment is poor, the device lacks the required camera set, or the scene does not produce stable person detections