blakeblackshear.frigate/frigate/track/norfair_tracker.py
gtsiam c0bd3b362c
Custom classes for Process and Metrics (#13950)
* Subclass Process for audio_process

* Introduce custom mp.Process subclass

In preparation to switch the multiprocessing startup method away from
"fork", we cannot rely on os.fork cloning the log state at fork time.
Instead, we have to set up logging before we run the business logic of
each process.

* Make camera_metrics into a class

* Make ptz_metrics into a class

* Fixed PtzMotionEstimator.ptz_metrics type annotation

* Removed pointless variables

* Do not start audio processor when no audio cameras are configured
2024-09-27 07:53:23 -05:00

409 lines
15 KiB
Python

import logging
import random
import string
import numpy as np
from norfair import (
Detection,
Drawable,
OptimizedKalmanFilterFactory,
Tracker,
draw_boxes,
)
from norfair.drawing.drawer import Drawer
from frigate.camera import PTZMetrics
from frigate.config import CameraConfig
from frigate.ptz.autotrack import PtzMotionEstimator
from frigate.track import ObjectTracker
from frigate.util.image import intersection_over_union
from frigate.util.object import average_boxes, median_of_boxes
logger = logging.getLogger(__name__)
THRESHOLD_KNOWN_ACTIVE_IOU = 0.2
THRESHOLD_STATIONARY_CHECK_IOU = 0.6
THRESHOLD_ACTIVE_CHECK_IOU = 0.9
MAX_STATIONARY_HISTORY = 10
# Normalizes distance from estimate relative to object size
# Other ideas:
# - if estimates are inaccurate for first N detections, compare with last_detection (may be fine)
# - could be variable based on time since last_detection
# - include estimated velocity in the distance (car driving by of a parked car)
# - include some visual similarity factor in the distance for occlusions
def distance(detection: np.array, estimate: np.array) -> float:
# ultimately, this should try and estimate distance in 3-dimensional space
# consider change in location, width, and height
estimate_dim = np.diff(estimate, axis=0).flatten()
detection_dim = np.diff(detection, axis=0).flatten()
# get bottom center positions
detection_position = np.array(
[np.average(detection[:, 0]), np.max(detection[:, 1])]
)
estimate_position = np.array([np.average(estimate[:, 0]), np.max(estimate[:, 1])])
distance = (detection_position - estimate_position).astype(float)
# change in x relative to w
distance[0] /= estimate_dim[0]
# change in y relative to h
distance[1] /= estimate_dim[1]
# get ratio of widths and heights
# normalize to 1
widths = np.sort([estimate_dim[0], detection_dim[0]])
heights = np.sort([estimate_dim[1], detection_dim[1]])
width_ratio = widths[1] / widths[0] - 1.0
height_ratio = heights[1] / heights[0] - 1.0
# change vector is relative x,y change and w,h ratio
change = np.append(distance, np.array([width_ratio, height_ratio]))
# calculate euclidean distance of the change vector
return np.linalg.norm(change)
def frigate_distance(detection: Detection, tracked_object) -> float:
return distance(detection.points, tracked_object.estimate)
class NorfairTracker(ObjectTracker):
def __init__(
self,
config: CameraConfig,
ptz_metrics: PTZMetrics,
):
self.tracked_objects = {}
self.untracked_object_boxes: list[list[int]] = []
self.disappeared = {}
self.positions = {}
self.stationary_box_history: dict[str, list[list[int, int, int, int]]] = {}
self.camera_config = config
self.detect_config = config.detect
self.ptz_metrics = ptz_metrics
self.ptz_motion_estimator = {}
self.camera_name = config.name
self.track_id_map = {}
# TODO: could also initialize a tracker per object class if there
# was a good reason to have different distance calculations
self.tracker = Tracker(
distance_function=frigate_distance,
distance_threshold=2.5,
initialization_delay=self.detect_config.min_initialized,
hit_counter_max=self.detect_config.max_disappeared,
# use default filter factory with custom values
# R is the multiplier for the sensor measurement noise matrix, default of 4.0
# lowering R means that we trust the position of the bounding boxes more
# testing shows that the prediction was being relied on a bit too much
# TODO: could use different kalman filter values along with
# the different tracker per object class
filter_factory=OptimizedKalmanFilterFactory(R=3.4),
)
if self.ptz_metrics.autotracker_enabled.value:
self.ptz_motion_estimator = PtzMotionEstimator(
self.camera_config, self.ptz_metrics
)
def register(self, track_id, obj):
rand_id = "".join(random.choices(string.ascii_lowercase + string.digits, k=6))
id = f"{obj['frame_time']}-{rand_id}"
self.track_id_map[track_id] = id
obj["id"] = id
obj["start_time"] = obj["frame_time"]
obj["motionless_count"] = 0
obj["position_changes"] = 0
obj["score_history"] = [
p.data["score"]
for p in next(
(o for o in self.tracker.tracked_objects if o.global_id == track_id)
).past_detections
]
self.tracked_objects[id] = obj
self.disappeared[id] = 0
self.positions[id] = {
"xmins": [],
"ymins": [],
"xmaxs": [],
"ymaxs": [],
"xmin": 0,
"ymin": 0,
"xmax": self.detect_config.width,
"ymax": self.detect_config.height,
}
self.stationary_box_history[id] = []
def deregister(self, id, track_id):
del self.tracked_objects[id]
del self.disappeared[id]
self.tracker.tracked_objects = [
o for o in self.tracker.tracked_objects if o.global_id != track_id
]
del self.track_id_map[track_id]
# tracks the current position of the object based on the last N bounding boxes
# returns False if the object has moved outside its previous position
def update_position(self, id: str, box: list[int, int, int, int], stationary: bool):
xmin, ymin, xmax, ymax = box
position = self.positions[id]
self.stationary_box_history[id].append(box)
if len(self.stationary_box_history[id]) > MAX_STATIONARY_HISTORY:
self.stationary_box_history[id] = self.stationary_box_history[id][
-MAX_STATIONARY_HISTORY:
]
avg_iou = intersection_over_union(
box, average_boxes(self.stationary_box_history[id])
)
# object has minimal or zero iou
# assume object is active
if avg_iou < THRESHOLD_KNOWN_ACTIVE_IOU:
self.positions[id] = {
"xmins": [xmin],
"ymins": [ymin],
"xmaxs": [xmax],
"ymaxs": [ymax],
"xmin": xmin,
"ymin": ymin,
"xmax": xmax,
"ymax": ymax,
}
return False
threshold = (
THRESHOLD_STATIONARY_CHECK_IOU if stationary else THRESHOLD_ACTIVE_CHECK_IOU
)
# object has iou below threshold, check median to reduce outliers
if avg_iou < threshold:
median_iou = intersection_over_union(
(
position["xmin"],
position["ymin"],
position["xmax"],
position["ymax"],
),
median_of_boxes(self.stationary_box_history[id]),
)
# if the median iou drops below the threshold
# assume object is no longer stationary
if median_iou < threshold:
self.positions[id] = {
"xmins": [xmin],
"ymins": [ymin],
"xmaxs": [xmax],
"ymaxs": [ymax],
"xmin": xmin,
"ymin": ymin,
"xmax": xmax,
"ymax": ymax,
}
return False
# if there are less than 10 entries for the position, add the bounding box
# and recompute the position box
if len(position["xmins"]) < 10:
position["xmins"].append(xmin)
position["ymins"].append(ymin)
position["xmaxs"].append(xmax)
position["ymaxs"].append(ymax)
# by using percentiles here, we hopefully remove outliers
position["xmin"] = np.percentile(position["xmins"], 15)
position["ymin"] = np.percentile(position["ymins"], 15)
position["xmax"] = np.percentile(position["xmaxs"], 85)
position["ymax"] = np.percentile(position["ymaxs"], 85)
return True
def is_expired(self, id):
obj = self.tracked_objects[id]
# get the max frames for this label type or the default
max_frames = self.detect_config.stationary.max_frames.objects.get(
obj["label"], self.detect_config.stationary.max_frames.default
)
# if there is no max_frames for this label type, continue
if max_frames is None:
return False
# if the object has exceeded the max_frames setting, deregister
if (
obj["motionless_count"] - self.detect_config.stationary.threshold
> max_frames
):
return True
return False
def update(self, track_id, obj):
id = self.track_id_map[track_id]
self.disappeared[id] = 0
stationary = (
self.tracked_objects[id]["motionless_count"]
>= self.detect_config.stationary.threshold
)
# update the motionless count if the object has not moved to a new position
if self.update_position(id, obj["box"], stationary):
self.tracked_objects[id]["motionless_count"] += 1
if self.is_expired(id):
self.deregister(id, track_id)
return
else:
# register the first position change and then only increment if
# the object was previously stationary
if (
self.tracked_objects[id]["position_changes"] == 0
or self.tracked_objects[id]["motionless_count"]
>= self.detect_config.stationary.threshold
):
self.tracked_objects[id]["position_changes"] += 1
self.tracked_objects[id]["motionless_count"] = 0
self.stationary_box_history[id] = []
self.tracked_objects[id].update(obj)
def update_frame_times(self, frame_time):
# if the object was there in the last frame, assume it's still there
detections = [
(
obj["label"],
obj["score"],
obj["box"],
obj["area"],
obj["ratio"],
obj["region"],
)
for id, obj in self.tracked_objects.items()
if self.disappeared[id] == 0
]
self.match_and_update(frame_time, detections=detections)
def match_and_update(self, frame_time, detections):
norfair_detections = []
for obj in detections:
# centroid is used for other things downstream
centroid_x = int((obj[2][0] + obj[2][2]) / 2.0)
centroid_y = int((obj[2][1] + obj[2][3]) / 2.0)
# track based on top,left and bottom,right corners instead of centroid
points = np.array([[obj[2][0], obj[2][1]], [obj[2][2], obj[2][3]]])
norfair_detections.append(
Detection(
points=points,
label=obj[0],
data={
"label": obj[0],
"score": obj[1],
"box": obj[2],
"area": obj[3],
"ratio": obj[4],
"region": obj[5],
"frame_time": frame_time,
"centroid": (centroid_x, centroid_y),
},
)
)
coord_transformations = None
if self.ptz_metrics.autotracker_enabled.value:
# we must have been enabled by mqtt, so set up the estimator
if not self.ptz_motion_estimator:
self.ptz_motion_estimator = PtzMotionEstimator(
self.camera_config, self.ptz_metrics
)
coord_transformations = self.ptz_motion_estimator.motion_estimator(
detections, frame_time, self.camera_name
)
tracked_objects = self.tracker.update(
detections=norfair_detections, coord_transformations=coord_transformations
)
# update or create new tracks
active_ids = []
for t in tracked_objects:
estimate = tuple(t.estimate.flatten().astype(int))
# keep the estimate within the bounds of the image
estimate = (
max(0, estimate[0]),
max(0, estimate[1]),
min(self.detect_config.width - 1, estimate[2]),
min(self.detect_config.height - 1, estimate[3]),
)
obj = {
**t.last_detection.data,
"estimate": estimate,
"estimate_velocity": t.estimate_velocity,
}
active_ids.append(t.global_id)
if t.global_id not in self.track_id_map:
self.register(t.global_id, obj)
# if there wasn't a detection in this frame, increment disappeared
elif t.last_detection.data["frame_time"] != frame_time:
id = self.track_id_map[t.global_id]
self.disappeared[id] += 1
# sometimes the estimate gets way off
# only update if the upper left corner is actually upper left
if estimate[0] < estimate[2] and estimate[1] < estimate[3]:
self.tracked_objects[id]["estimate"] = obj["estimate"]
# else update it
else:
self.update(t.global_id, obj)
# clear expired tracks
expired_ids = [k for k in self.track_id_map.keys() if k not in active_ids]
for e_id in expired_ids:
self.deregister(self.track_id_map[e_id], e_id)
# update list of object boxes that don't have a tracked object yet
tracked_object_boxes = [obj["box"] for obj in self.tracked_objects.values()]
self.untracked_object_boxes = [
o[2] for o in detections if o[2] not in tracked_object_boxes
]
def debug_draw(self, frame, frame_time):
active_detections = [
Drawable(id=obj.id, points=obj.last_detection.points, label=obj.label)
for obj in self.tracker.tracked_objects
if obj.last_detection.data["frame_time"] == frame_time
]
missing_detections = [
Drawable(id=obj.id, points=obj.last_detection.points, label=obj.label)
for obj in self.tracker.tracked_objects
if obj.last_detection.data["frame_time"] != frame_time
]
# draw the estimated bounding box
draw_boxes(frame, self.tracker.tracked_objects, color="green", draw_ids=True)
# draw the detections that were detected in the current frame
draw_boxes(frame, active_detections, color="blue", draw_ids=True)
# draw the detections that are missing in the current frame
draw_boxes(frame, missing_detections, color="red", draw_ids=True)
# draw the distance calculation for the last detection
# estimate vs detection
for obj in self.tracker.tracked_objects:
ld = obj.last_detection
# bottom right
text_anchor = (
ld.points[1, 0],
ld.points[1, 1],
)
frame = Drawer.text(
frame,
f"{obj.id}: {str(obj.last_distance)}",
position=text_anchor,
size=None,
color=(255, 0, 0),
thickness=None,
)