"""
SORT: A Simple, Online and Realtime Tracker
Copyright (C) 2016 Alex Bewley alex@dynamicdetection.com
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
"""
from __future__ import print_function
from numba import jit
from typing import List
import numpy as np
from sklearn.utils.linear_assignment_ import linear_assignment
from filterpy.kalman import KalmanFilter
import logging
@jit
[docs]def iou(bb_test, bb_gt):
"""
Computes IUO between two bboxes in the form [x1,y1,x2,y2]
"""
xx1 = np.maximum(bb_test[0], bb_gt[0])
yy1 = np.maximum(bb_test[1], bb_gt[1])
xx2 = np.minimum(bb_test[2], bb_gt[2])
yy2 = np.minimum(bb_test[3], bb_gt[3])
w = np.maximum(0., xx2 - xx1)
h = np.maximum(0., yy2 - yy1)
wh = w * h
o = wh / ((bb_test[2] - bb_test[0]) * (bb_test[3] - bb_test[1])
+ (bb_gt[2] - bb_gt[0]) * (bb_gt[3] - bb_gt[1]) - wh)
return o
# @jit(nopython=True)
@jit
[docs]def convert_bbox_to_z(bbox):
"""
Takes a bounding box in the form [x1,y1,x2,y2] and returns z in the form
[x,y,s,r] where x,y is the centre of the box and s is the scale/area and r is
the aspect ratio
"""
w = bbox[2] - bbox[0]
h = bbox[3] - bbox[1]
x = bbox[0] + w / 2.
y = bbox[1] + h / 2.
s = w * h # scale is just area
r = w / float(h)
return np.array([x, y, s, r]).reshape((4, 1))
# @jit(nopython=True)
@jit
[docs]def convert_x_to_bbox(x, score=None):
"""
Takes a bounding box in the centre form [x,y,s,r] and returns it in the form
[x1,y1,x2,y2] where x1,y1 is the top left and x2,y2 is the bottom right
"""
w = np.sqrt(x[2] * x[3])
h = x[2] / w
if score is None:
return np.array([x[0] - w / 2., x[1] - h / 2., x[0] + w / 2., x[1] + h / 2.]).reshape((1, 4))
else:
return np.array([x[0] - w / 2., x[1] - h / 2., x[0] + w / 2., x[1] + h / 2., score]).reshape((1, 5))
[docs]class KalmanBoxTracker(object):
"""
This class represents the internel state of individual tracked objects observed as bbox.
"""
def __init__(self, bbox, window_size=None):
"""
Initialises a tracker using initial bounding box.
"""
# define constant velocity model
self.kf = KalmanFilter(dim_x=7, dim_z=4)
self.kf.F = np.array(
[[1, 0, 0, 0, 1, 0, 0], [0, 1, 0, 0, 0, 1, 0], [0, 0, 1, 0, 0, 0, 1], [0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 1, 0], [0, 0, 0, 0, 0, 0, 1]])
self.kf.H = np.array(
[[1, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0]])
self.kf.R[2:, 2:] *= 10.
self.kf.P[4:, 4:] *= 1000. # give high uncertainty to the unobservable initial velocities
self.kf.P *= 10.
self.kf.Q[-1, -1] *= 0.01
self.kf.Q[4:, 4:] *= 0.01
self.kf.x[:4] = convert_bbox_to_z(bbox)
self.time_since_update = 0
self.id = KalmanBoxTracker.count
KalmanBoxTracker.count += 1
self.history = []
self.hits = 0
self.hit_streak = 0
self.seen_in_window = []
self.window_size = window_size
if self.window_size:
self.seen_in_window = [0] * self.window_size
self.seen_in_window[-1] = 1
self.age = 0
self.extra_info = bbox[4:]
[docs] def update(self, bbox):
"""
Updates the state vector with observed bbox.
"""
self.time_since_update = 0
self.history = []
self.hits += 1
self.hit_streak += 1
if self.window_size:
self.seen_in_window[-1] = 1
self.kf.update(convert_bbox_to_z(bbox))
self.extra_info = bbox[4:]
[docs] def predict(self):
"""
Advances the state vector and returns the predicted bounding box estimate.
"""
if (self.kf.x[6] + self.kf.x[2]) <= 0:
self.kf.x[6] *= 0.0
self.kf.predict()
self.age += 1
if self.window_size:
self.seen_in_window.append(0)
self.seen_in_window.pop(0)
if self.time_since_update > 0:
self.hit_streak = 0
self.time_since_update += 1
self.history.append(convert_x_to_bbox(self.kf.x))
return self.history[-1]
[docs] def get_state(self):
"""
Returns the current bounding box estimate.
"""
return convert_x_to_bbox(self.kf.x)
[docs]def associate_detections_to_trackers(detections, trackers, iou_threshold=0.3):
"""
Assigns detections to tracked object (both represented as bounding boxes)
Returns 3 lists of matches, unmatched_detections and unmatched_trackers
"""
if len(trackers) == 0:
return np.empty((0, 2), dtype=int), np.arange(len(detections)), np.empty((0, 5), dtype=int)
iou_matrix = np.zeros((len(detections), len(trackers)), dtype=np.float32)
for d, det in enumerate(detections):
for t, trk in enumerate(trackers):
iou_matrix[d, t] = iou(det, trk)
matched_indices = linear_assignment(-iou_matrix)
unmatched_detections = []
for d, det in enumerate(detections):
if d not in matched_indices[:, 0]:
unmatched_detections.append(d)
unmatched_trackers = []
for t, trk in enumerate(trackers):
if t not in matched_indices[:, 1]:
unmatched_trackers.append(t)
# filter out matched with low IOU
matches = []
for m in matched_indices:
if iou_matrix[m[0], m[1]] < iou_threshold:
unmatched_detections.append(m[0])
unmatched_trackers.append(m[1])
else:
matches.append(m.reshape(1, 2))
if len(matches) == 0:
matches = np.empty((0, 2), dtype=int)
else:
matches = np.concatenate(matches, axis=0)
return matches, np.array(unmatched_detections), np.array(unmatched_trackers)
[docs]class Sort:
def __init__(self, max_age: int = 1, min_hits: int = None, window_size: int = None, percent_seen: float = None,
verbose: bool = False):
"""
Args:
max_age: number of times track is missed before stops looking for ot
min_hits: number of times a track is seen before returns it in the object update method
window_size: size of frame window for non-continuous track
percent_seen: percent of frames in window_size a track is seen before it is returned in the object update
verbose: if True, set DEBUG as logger level
"""
super(Sort, self).__init__()
self.max_age = max_age
self.min_hits = min_hits
self.window_size = window_size
self.percent_seen = percent_seen
self.trackers: List[KalmanBoxTracker] = []
self.frame_count = 0
self.logger = logging.getLogger(__name__)
if verbose:
self.logger.setLevel(logging.DEBUG)
self.logger.debug(f"SORT tracker initialized with 'max age':{max_age}, 'min hits':{min_hits}, "
f"'window_size':{window_size}, 'percent_seen':{percent_seen}")
if not (self.min_hits is None) ^ (self.window_size is None):
raise ValueError("Exactly one of `min_hits`, or `window_size` "
"arguments must be provided.")
if self.percent_seen is None and self.window_size is not None:
raise ValueError("If `window_size` is provided, then `percent_seen` "
"should be also provided.")
[docs] def reset(self):
"""
reset the tracker, the same functionality as initializing a new Sort object
"""
self.trackers = []
self.frame_count = 0
self.logger.debug("SORT tracker reset")
[docs] def update(self, dets: np.array):
"""
Params:
dets - a numpy array of detections in the format [[x1,y1,x2,y2,score],[x1,y1,x2,y2,score],...]
Requires: this method must be called once for each frame even with empty detections.
Returns the a similar array, where the last column is the object ID.
NOTE: The number of objects returned may differ from the number of detections provided.
"""
self.frame_count += 1
# get predicted locations from existing trackers.
trks = np.zeros((len(self.trackers), 5))
to_del = []
ret = []
for t, trk in enumerate(trks):
pos = self.trackers[t].predict()[0]
trk[:] = [pos[0], pos[1], pos[2], pos[3], 0]
if np.any(np.isnan(pos)):
to_del.append(t)
trks = np.ma.compress_rows(np.ma.masked_invalid(trks))
for t in reversed(to_del):
self.trackers.pop(t)
matched, unmatched_dets, unmatched_trks = associate_detections_to_trackers(dets, trks)
# update matched trackers with assigned detections
for t, trk in enumerate(self.trackers):
if t not in unmatched_trks:
d = matched[np.where(matched[:, 1] == t)[0], 0]
trk.update(dets[d, :][0])
# create and initialise new trackers for unmatched detections
for i in unmatched_dets:
trk = KalmanBoxTracker(dets[i, :], window_size=self.window_size)
self.trackers.append(trk)
i = len(self.trackers)
for trk in reversed(self.trackers):
d = trk.get_state()[0]
if self.min_hits:
seen_enough = trk.hit_streak >= self.min_hits
min_frames = self.min_hits
else:
seen_enough = (np.mean(trk.seen_in_window) >= self.percent_seen)
min_frames = self.window_size
if (trk.time_since_update < 1) and (seen_enough or self.frame_count <= min_frames):
# trk.id + 1 as MOT benchmark requires positive
ret.append(np.concatenate((d, trk.extra_info, [trk.id + 1])).reshape(1, -1))
i -= 1
# remove dead tracklet
if trk.time_since_update > self.max_age:
self.trackers.pop(i)
self.logger.debug(f"Update: unmatched detections-{len(unmatched_dets)}; unmatched tracks-{len(unmatched_trks)}"
f"; deleted tracks-{len(to_del)}; matched-{len(matched)}; returned-{len(ret)}")
if len(ret) > 0:
return np.concatenate(ret)
# return np.empty((0, 5))
return None