Source code for inaFaceAnalyzer.face_tracking

#!/usr/bin/env python
# encoding: utf-8

# The MIT License

# Copyright (c) 2021 Ina (David Doukhan & Zohra Rezgui- http://www.ina.fr/)

# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:

# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.

# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.

from typing import NamedTuple
import dlib
import numpy as np
from .rect import Rect
from .opencv_utils import disp_frame_shapes


[docs]class TrackDetection(NamedTuple): """ Atomic element returned by face tracking and detection classes """ #: bounding box bbox : Rect #: detected face numerical identifier face_id : int #: face detection confidence detect_conf : float #: tracking confidence track_conf : float
def _matrix_argmax(m): x = np.argmax(m) dim = m.shape[1] return x // dim, x % dim class Tracker: def __init__(self, frame, bb, detect_conf): self.t = dlib.correlation_tracker() if not isinstance(bb, dlib.drectangle): bb = dlib.drectangle(*bb) self.t.start_track(frame, bb) self.fshape = frame.shape self.detect_conf = detect_conf self.track_conf = None def update(self, frame, verbose=False): update_val = self.t.update(frame) fh, fw, _ = self.fshape e = self.t.get_position() x1, y1, x2, y2 = pos = Rect.from_dlib(e) if verbose: print('update', pos, update_val) disp_frame_shapes(frame, [pos]) if not ((x2 > 0) and (x1 < fw) and (y2 > 0) and (y1 < fh)): update_val = -1 self.detect_conf = None self.track_conf = update_val return update_val def update_from_detection(self, frame, dtc, verbose=False): update_val = self.t.update(frame, dtc.bbox.to_dlibFloat()) if verbose: e = self.t.get_position() print('dest', dtc.bbox, 'new position', e) disp_frame_shapes(frame, [Rect.from_dlib(e), dtc.bbox]) self.t.start_track(frame, dtc.bbox.to_dlibFloat()) self.track_conf = update_val self.detect_conf = dtc.detect_conf return update_val class TrackerDetector: # confidence is estimated as the peak to side-lobe ration returned # on tracker's update min_confidence = 7 # output labels output_type = TrackDetection #out_names = ['bb', 'face_id', 'face_detect_conf', 'tracking_conf'] def __init__(self, detector, detection_period): # dictionnary of tracked objects self.d = {} # number of instantiated trackers # used to provide a numeric identifier to each new Tracker instance self.nb_tracker = 0 # a face/object detection instance # could be any face detector class defined in face_detector.py self.detector = detector # Detection will be performed once every 'detection_period' frames # ex: if detection_period is set to 5, detection will be performed # for 1 frame, and tracking for the 4 remaining frames # if detection_period is set to 1, the detection will be performed on # every frame, and tracking will also be performed to know the detected # faces belong to the same person self.detection_period = detection_period # count the amount of processed frames # used to switch between detection and tracking every detection_period self.iframe = 0 def update_trackers(self, frame, verbose = False): if verbose: print('update trackers') # if tracked element is lost, remove tracker for fid in list(self.d): if self.d[fid].update(frame, verbose) < self.min_confidence: del self.d[fid] if verbose: print('deleting tracker', fid) def update_from_detection(self, frame, ldetections, verbose): if verbose: print('update from detection') lkeys = list(self.d.keys()) # compute intersection over union matrix[#tracker, #detected bounding box] # between tracker positions and detected bounding boxes ioumat = np.ones((len(lkeys), len(ldetections))) * -1 for i, k in enumerate(lkeys): trackpos = Rect.from_dlib(self.d[k].t.get_position()) for j, detection in enumerate(ldetections): ioumat[i, j] = detection.bbox.iou(trackpos) # while matrix not empty and IOU > 70% while np.prod(ioumat.shape): # find the largest intersection over union itracker, idetection = am = _matrix_argmax(ioumat) if ioumat[am] <= 0.7: break # update closest bounding box and trackers and remove them from matrix track_score = self.d[lkeys[itracker]].update_from_detection(frame, ldetections[idetection], verbose) ioumat = np.delete(ioumat, itracker, axis = 0) k = lkeys.pop(itracker) # if close bounding box and tracker do not match, delete tracker if track_score < self.min_confidence: del self.d[k] else: # if bounding box and detected face match, remove # the detection from the set ldetections.pop(idetection) ioumat = np.delete(ioumat, idetection, axis = 1) # remove trackers that do not match any detected box for k in lkeys: del self.d[k] # add new trackers corresponding to detected faces that did not match # any existing tracker for dtc in ldetections: self.d[self.nb_tracker] = Tracker(frame, dtc.bbox, dtc.detect_conf) self.nb_tracker += 1 def __call__(self, frame, verbose=False): if self.iframe % self.detection_period == 0: self.update_from_detection(frame, self.detector(frame, verbose), verbose) else: self.update_trackers(frame, verbose) lret = [] for faceid in self.d: t = self.d[faceid] bb = t.t.get_position() lret.append(TrackDetection(Rect.from_dlib(bb), faceid, t.detect_conf, t.track_conf)) self.iframe += 1 return lret