Configuration

Configuration

Runtime behaviour is driven by config/default.yaml. Copy or edit it to tune the pipeline. Most values also have sensible in-app defaults, and many are adjustable live from the UI. This page groups every top-level section.

Environment overrides for the model tier and devices are listed on the Models page (OVERSEER_STT_MODEL, OVERSEER_CLIP_DEVICE, ...).

Core

Section	Keys	Notes
app	log_level, log_dir, data_dir	Base paths. data_dir (default data) holds the SQLite db, snapshots, clips, media.
camera	connect_timeout, read_timeout, reconnect_min_delay, reconnect_max_delay, buffer_size, freeze_timeout	Capture and reconnect behaviour.
media	max_height, max_concurrent_downloads	Cap downloaded (YouTube) video resolution; serialize downloads.
thumbs	max_workers	Parallel map-preview relays.

Detection & perception

Section	Keys	Notes
detectors.motion	enabled, min_area, event_threshold_percent	MOG2 motion.
detectors.yolo	model, confidence, person_confidence, frame_interval, imgsz, slice, slice_overlap	The main detector. person_confidence uses a lower floor for people (awkward poses). slice enables tiled inference for tiny distant objects.
detectors.{person,vehicle,animal,accessory,weapon}	enabled	Per-class gating, applied at the detector so disabling a class removes it from the whole downstream budget.
analyzers	frame_interval, daynight, fog, rain, wind, weather	Scene condition analysers, each with onset/clear thresholds.
pose	pose estimation settings	Body pose for posture and behaviour.
trajectory	buffer_size, expire_seconds, running_speed, stopped_seconds, stopped_eps, uturn_window, uturn_angle	Movement events: running, stopped, U-turn.
speed	class_heights, default_height, ...	Rough km/h estimate from apparent size and the ground plane.
egomotion	ego-motion flow settings	Compensates a moving / dashcam feed so speeds read against the ground.
vehicle	make.model, make.min_conf, make.min_margin, make.min_area, make.live_interval, bodytype.*	Brand + body-type classifier gating and throttle.
gait	gait descriptor settings	Soft biometric for cross-camera Re-ID.

Identity & roster

Section	Keys	Notes
roster	enabled, detector, confidence, imgsz, interval, dedup_threshold, persist, persist_threshold, fullres_interval, clips, clip_frames, clip_fps, watch_cooldown, auto_merge.*	The background harvester that fills the roster from all cameras, deduplicates by ReID embedding, persists long-term subjects, and auto-merges very-confident duplicates with safety rails.
match	appearance-search backends	Which ReID encoders back forensic and cross-camera matching.
reconstruct	multi-frame reconstruction settings	Super-resolved subject reconstruction from many sightings.

Analytics, alerts & spatial

Section	Keys	Notes
events	throttle_seconds, snapshot_on_event	Event de-duplication and snapshotting.
zones	zone defaults	Polygon and line zones.
alerts	alert engine settings	Rule engine + cooldowns.
threat	correlation-based threat scoring	Fuses co-occurring events so dangerous combinations escalate.
anomaly	anomaly detection settings	Learned unusual-activity flags.
suggestions	min_coverage, ...	Smart alert + zone suggestions (auto zone detection from foot-point density).
spatial	spatial 3D build settings	Depth fusion, point cloud, mesh, background completion.
cases	case defaults	Investigation cases.
heatmap, statistics, retention, recording, timelapse	outputs and housekeeping	Crowd density, rollup counts, data retention, recording modes, timelapse.
objects	object-of-interest tracker settings	Track an arbitrary boxed object (OOI).

Server & AI

Section	Keys	Notes
overseer, rest_api	server bind / API settings	Host and port.
forensic	forensic search settings	Attribute parsing and confidence floors.
ai	LLM provider config	The AI Operator planner, chat, VQA. Configure a provider in Settings; see AI Operator.

This table names the sections and their most relevant keys. The authoritative, commented reference is config/default.yaml in the repository, which explains the reasoning behind each default inline.