tiny-trigger / tiny_trigger /automation.py
Javier Montalvo
Decouple tracking from detection; size-relative motion; UI tuning
114ea19
Raw
History Blame Contribute Delete
18.6 kB
from __future__ import annotations
import json
from collections import defaultdict
from math import hypot
from statistics import median
from typing import Any, Literal
from pydantic import AliasChoices, BaseModel, ConfigDict, Field, ValidationError, model_validator
from .models import ActionEvent, Detection, FrameSample
def _label(value: str) -> str:
return value.strip().lower()
def _dedupe_labels(labels: list[str]) -> list[str]:
seen: set[str] = set()
result: list[str] = []
for label in labels:
normalized = " ".join(label.strip().split())
key = normalized.lower()
if normalized and key not in seen:
seen.add(key)
result.append(normalized)
return result
class PresentCondition(BaseModel):
label: str
min_count: int = Field(default=1, ge=1)
class CountCondition(BaseModel):
label: str
minimum: int = Field(default=1, ge=1, validation_alias=AliasChoices("min", "minimum", "min_count"))
class NearCondition(BaseModel):
a: str
b: str
max_gap_percent: float = Field(
default=16.0,
ge=0.0,
validation_alias=AliasChoices("max_gap_percent", "max_distance"),
)
@model_validator(mode="before")
@classmethod
def migrate_max_distance(cls, data: Any) -> Any:
if isinstance(data, dict) and "max_distance" in data and "max_gap_percent" not in data:
value = data["max_distance"]
if isinstance(value, int | float) and value <= 1:
return {**data, "max_gap_percent": value * 100.0}
return data
class FarCondition(BaseModel):
a: str
b: str
min_gap_percent: float = Field(default=25.0, ge=0.0)
class MovingCondition(BaseModel):
label: str
min_displacement_ratio: float = Field(default=0.15, ge=0.0)
window_frames: int = Field(default=3, ge=3)
max_missing_frames: int = Field(default=1, ge=0)
@model_validator(mode="before")
@classmethod
def migrate_short_window(cls, data: Any) -> Any:
if isinstance(data, dict) and data.get("window_frames") is not None:
try:
window_frames = int(data["window_frames"])
except (TypeError, ValueError):
return data
if window_frames < 3:
return {**data, "window_frames": 3}
return data
class CooldownCondition(BaseModel):
key: str | None = None
seconds: float | None = Field(default=None, ge=0.0)
minutes: float | None = Field(default=None, ge=0.0)
@property
def duration_seconds(self) -> float:
if self.seconds is not None:
return self.seconds
if self.minutes is not None:
return self.minutes * 60.0
return 0.0
class ConditionBlock(BaseModel):
present: PresentCondition | None = None
count: CountCondition | None = None
near: NearCondition | None = None
far: FarCondition | None = None
moving: MovingCondition | None = None
cooldown: CooldownCondition | None = None
@model_validator(mode="after")
def exactly_one_condition(self) -> "ConditionBlock":
selected = [self.present, self.count, self.near, self.far, self.moving, self.cooldown]
if sum(item is not None for item in selected) != 1:
raise ValueError("Each condition block must contain exactly one condition.")
return self
class WhenClause(BaseModel):
model_config = ConfigDict(populate_by_name=True)
all_conditions: list[ConditionBlock] = Field(default_factory=list, alias="all")
any_conditions: list[ConditionBlock] = Field(default_factory=list, alias="any")
@model_validator(mode="after")
def has_conditions(self) -> "WhenClause":
if not self.all_conditions and not self.any_conditions:
raise ValueError("A rule needs at least one condition.")
return self
class ActionSpec(BaseModel):
type: Literal["simulate", "webhook"] = "simulate"
name: str
url: str | None = None
payload: dict[str, Any] = Field(default_factory=dict)
class TriggerClause(BaseModel):
on: Literal["while", "enter", "exit", "change"] = "enter"
class ActionSet(BaseModel):
model_config = ConfigDict(populate_by_name=True)
enter: list[ActionSpec] = Field(default_factory=list)
exit: list[ActionSpec] = Field(default_factory=list)
while_actions: list[ActionSpec] = Field(default_factory=list, alias="while")
class GateClause(BaseModel):
enabled: bool = True
cooldown: CooldownCondition | None = None
class AutomationRule(BaseModel):
name: str
when: WhenClause
trigger: TriggerClause = Field(default_factory=TriggerClause)
gate: GateClause = Field(default_factory=GateClause)
then: list[ActionSpec] | ActionSet
@model_validator(mode="after")
def has_actions_for_trigger(self) -> "AutomationRule":
if isinstance(self.then, list):
if not self.then:
raise ValueError("A rule needs at least one action.")
if self.trigger.on == "change":
raise ValueError("change triggers need then.enter/then.exit actions.")
return self
actions = _actions_for_trigger(self.then, self.trigger.on)
if not actions:
raise ValueError(f"A rule with trigger.on={self.trigger.on!r} needs matching actions.")
return self
class AutomationDocument(BaseModel):
rules: list[AutomationRule] = Field(default_factory=list)
def load_automation_text(text: str) -> AutomationDocument:
"""Load a JSON/YAML automation document and validate it."""
raw = text.strip()
if not raw:
raise ValueError("Rules are empty.")
try:
data = json.loads(raw)
except json.JSONDecodeError:
try:
import yaml
except ImportError as exc: # pragma: no cover - dependency guard
raise RuntimeError("Install PyYAML to load YAML automation rules.") from exc
data = yaml.safe_load(raw)
if isinstance(data, list):
data = {"rules": data}
if isinstance(data, dict) and "rules" not in data and "name" in data:
data = {"rules": [data]}
try:
return AutomationDocument.model_validate(data)
except ValidationError:
raise
except Exception as exc:
raise ValueError(f"Invalid automation document: {exc}") from exc
def automation_schema() -> dict[str, Any]:
return AutomationDocument.model_json_schema()
def rule_labels(rule: AutomationRule) -> list[str]:
labels: list[str] = []
for condition in [*rule.when.all_conditions, *rule.when.any_conditions]:
if condition.present:
labels.append(condition.present.label)
if condition.count:
labels.append(condition.count.label)
if condition.near:
labels.extend([condition.near.a, condition.near.b])
if condition.far:
labels.extend([condition.far.a, condition.far.b])
if condition.moving:
labels.append(condition.moving.label)
return _dedupe_labels(labels)
def document_labels(document: AutomationDocument, *, enabled_only: bool = True) -> list[str]:
labels: list[str] = []
for rule in document.rules:
if enabled_only and not rule.gate.enabled:
continue
labels.extend(rule_labels(rule))
return _dedupe_labels(labels)
class RuleEngine:
def __init__(
self,
rules: list[AutomationRule],
last_fired: dict[str, float] | None = None,
last_matched: dict[str, bool] | None = None,
) -> None:
self.rules = rules
self.last_fired: dict[str, float] = dict(last_fired or {})
self.last_matched: dict[str, bool] = dict(last_matched or {})
self.track_history: dict[int, list[tuple[int, tuple[float, float], float]]] = {}
self.moving_track_last_seen: dict[tuple[str, int], int] = {}
def evaluate_frame(
self,
detections: list[Detection],
*,
frame_index: int,
timestamp_sec: float,
) -> list[ActionEvent]:
events: list[ActionEvent] = []
self._update_track_history(detections)
for rule in self.rules:
if not self._gate_allows(rule, timestamp_sec):
self.last_matched[rule.name] = False
continue
matched = self._rule_matches(rule, detections, frame_index, timestamp_sec)
previous = self.last_matched.get(rule.name, False)
edge = _trigger_edge(previous=previous, matched=matched)
self.last_matched[rule.name] = matched
actions = _actions_to_fire(rule, edge)
if not matched and not actions:
continue
self._mark_cooldowns(rule, timestamp_sec)
for action in actions:
payload = {
"rule": rule.name,
"action": action.name,
"trigger": edge,
"frame_index": frame_index,
"timestamp_sec": timestamp_sec,
"detections": [item.model_dump(mode="json") for item in detections],
}
payload.update(action.payload)
events.append(
ActionEvent(
rule=rule.name,
action=action.name,
type=action.type,
frame_index=frame_index,
timestamp_sec=timestamp_sec,
url=action.url,
payload=payload,
)
)
return events
def _rule_matches(
self,
rule: AutomationRule,
detections: list[Detection],
frame_index: int,
timestamp_sec: float,
) -> bool:
all_ok = all(
self._condition_matches(condition, detections, frame_index, rule.name, timestamp_sec)
for condition in rule.when.all_conditions
)
any_ok = True
if rule.when.any_conditions:
any_ok = any(
self._condition_matches(condition, detections, frame_index, rule.name, timestamp_sec)
for condition in rule.when.any_conditions
)
return all_ok and any_ok
def _gate_allows(self, rule: AutomationRule, timestamp_sec: float) -> bool:
if not rule.gate.enabled:
return False
if not rule.gate.cooldown:
return True
return self._cooldown_allows(rule.gate.cooldown, rule.name, timestamp_sec)
def _condition_matches(
self,
condition: ConditionBlock,
detections: list[Detection],
frame_index: int,
rule_name: str,
timestamp_sec: float,
) -> bool:
by_label = _group_by_label(detections)
if condition.present:
return len(by_label[_label(condition.present.label)]) >= condition.present.min_count
if condition.count:
return len(by_label[_label(condition.count.label)]) >= condition.count.minimum
if condition.near:
left = by_label[_label(condition.near.a)]
right = by_label[_label(condition.near.b)]
if not left or not right:
return False
return _min_box_gap_percent(left, right) <= condition.near.max_gap_percent
if condition.far:
left = by_label[_label(condition.far.a)]
right = by_label[_label(condition.far.b)]
if not left or not right:
return False
return _min_box_gap_percent(left, right) >= condition.far.min_gap_percent
if condition.moving:
return self._moving_matches(condition.moving, by_label, frame_index)
if condition.cooldown:
return self._cooldown_allows(condition.cooldown, rule_name, timestamp_sec)
return False
def _moving_matches(
self,
condition: MovingCondition,
by_label: dict[str, list[Detection]],
frame_index: int,
) -> bool:
label = _label(condition.label)
label_detections = by_label[label]
for detection in label_detections:
if detection.track_id is None:
continue
history = self.track_history.get(detection.track_id, [])
if len(history) < condition.window_frames:
continue
window = history[-condition.window_frames :]
# Smooth out box jitter by averaging each half of the window, then
# measure displacement relative to the object's own size (its median
# box diagonal) so the threshold is distance-invariant and a flickering
# box on a stationary object stays well under it.
half = condition.window_frames // 2
early = _mean_point([point for _frame, point, _size in window[:half]])
late = _mean_point([point for _frame, point, _size in window[-half:]])
size = median(size for _frame, _point, size in window)
if size <= 0:
continue
displacement = hypot(late[0] - early[0], late[1] - early[1])
if displacement / size >= condition.min_displacement_ratio:
self.moving_track_last_seen[(label, detection.track_id)] = detection.frame_index
return True
if label_detections:
return False
return any(
last_seen_frame <= frame_index
and frame_index - last_seen_frame <= condition.max_missing_frames
for (track_label, _track_id), last_seen_frame in self.moving_track_last_seen.items()
if track_label == label
)
def _update_track_history(self, detections: list[Detection]) -> None:
for detection in detections:
if detection.track_id is None:
continue
history = self.track_history.setdefault(detection.track_id, [])
history.append(
(detection.frame_index, _foot_point_percent(detection), _box_diagonal_percent(detection))
)
del history[:-10]
def _mark_cooldowns(self, rule: AutomationRule, timestamp_sec: float) -> None:
if rule.gate.cooldown:
self.last_fired[rule.gate.cooldown.key or rule.name] = timestamp_sec
for condition in [*rule.when.all_conditions, *rule.when.any_conditions]:
if condition.cooldown:
self.last_fired[condition.cooldown.key or rule.name] = timestamp_sec
def _cooldown_allows(self, cooldown: CooldownCondition, rule_name: str, timestamp_sec: float) -> bool:
key = cooldown.key or rule_name
last = self.last_fired.get(key)
return last is None or (timestamp_sec - last) >= cooldown.duration_seconds
def evaluate_video_detections(
rules: list[AutomationRule],
detections: list[Detection],
*,
frames: list[FrameSample] | None = None,
last_fired: dict[str, float] | None = None,
) -> tuple[list[ActionEvent], dict[str, float]]:
engine = RuleEngine(rules, last_fired=last_fired)
events: list[ActionEvent] = []
grouped: dict[tuple[int, float], list[Detection]] = defaultdict(list)
for detection in detections:
grouped[(detection.frame_index, detection.timestamp_sec)].append(detection)
if frames:
for frame in frames:
grouped.setdefault((frame.frame_index, frame.timestamp_sec), [])
for (frame_index, timestamp_sec), frame_detections in sorted(grouped.items()):
events.extend(
engine.evaluate_frame(
frame_detections,
frame_index=frame_index,
timestamp_sec=timestamp_sec,
)
)
return events, dict(engine.last_fired)
def _group_by_label(detections: list[Detection]) -> dict[str, list[Detection]]:
grouped: dict[str, list[Detection]] = defaultdict(list)
for detection in detections:
grouped[_label(detection.label)].append(detection)
return grouped
def _min_box_gap_percent(left: list[Detection], right: list[Detection]) -> float:
best = float("inf")
for left_detection in left:
for right_detection in right:
best = min(best, _box_gap_percent(left_detection, right_detection))
return best
def _box_gap_percent(left: Detection, right: Detection) -> float:
ax1, ay1, ax2, ay2 = left.bbox_xyxy_norm
bx1, by1, bx2, by2 = right.bbox_xyxy_norm
gap_x = max(0.0, max(bx1 - ax2, ax1 - bx2))
gap_y = max(0.0, max(by1 - ay2, ay1 - by2))
return max(gap_x, gap_y) * 100.0
def _foot_point_percent(detection: Detection) -> tuple[float, float]:
# Bottom-center ("foot point"): far steadier than the box center for a
# standing/grounded object, whose top edge flickers with arms/hair/occlusion.
x1, _y1, x2, y2 = detection.bbox_xyxy_norm
return ((x1 + x2) * 50.0, y2 * 100.0)
def _box_diagonal_percent(detection: Detection) -> float:
x1, y1, x2, y2 = detection.bbox_xyxy_norm
return hypot((x2 - x1) * 100.0, (y2 - y1) * 100.0)
def _mean_point(points: list[tuple[float, float]]) -> tuple[float, float]:
count = len(points)
return (sum(point[0] for point in points) / count, sum(point[1] for point in points) / count)
def _trigger_edge(*, previous: bool, matched: bool) -> Literal["enter", "exit", "while", "none"]:
if matched and not previous:
return "enter"
if not matched and previous:
return "exit"
if matched:
return "while"
return "none"
def _actions_to_fire(rule: AutomationRule, edge: str) -> list[ActionSpec]:
trigger = rule.trigger.on
if trigger == "while":
return _actions_for_trigger(rule.then, "while") if edge in {"enter", "while"} else []
if trigger == "enter":
return _actions_for_trigger(rule.then, "enter") if edge == "enter" else []
if trigger == "exit":
return _actions_for_trigger(rule.then, "exit") if edge == "exit" else []
if trigger == "change":
return _actions_for_trigger(rule.then, edge) if edge in {"enter", "exit"} else []
return []
def _actions_for_trigger(
actions: list[ActionSpec] | ActionSet,
trigger: Literal["while", "enter", "exit", "change"],
) -> list[ActionSpec]:
if isinstance(actions, list):
return actions if trigger in {"while", "enter", "exit"} else []
if trigger == "while":
return actions.while_actions or actions.enter
if trigger == "enter":
return actions.enter
if trigger == "exit":
return actions.exit
return [*actions.enter, *actions.exit]