基于Mxnet实现目标跟踪SMOT【附部分源码】

---

前言

本文基于Mxnet实现SMOT，多目标跟踪算法，这里先使用COCO数据集的目标检测检测对象，然后在使用SMOT进行目标跟踪

---

一、多目标定位的类的实现

class GluonSSDMultiClassTracktor(BaseAnchorBasedTracktor):
    def __init__(self, gpu_id=0, detector_thresh=0.5, model_name="", use_pretrained=False, param_path="", data_shape=512,classes=[]):
        self.detector = GeneralDetector(gpu_id, aspect_ratio=1., data_shape=data_shape, model_name=model_name, use_pretrained=use_pretrained, param_path=param_path,classes = classes)
        self._anchor_tensor = None
        self._detector_thresh = detector_thresh
        self._ctx = mx.gpu(gpu_id)
        self._dummy_ti = mx.nd.array([[0]], ctx=self._ctx)

    def anchors(self):
        if self.detector.anchor_tensor is None:
            raise ValueError("anchor not initialized yet")
        return self.detector.anchor_tensor

    def clean_up(self):
        pass

    def prepare_for_frame(self, frame):
        return None

    @timeit
    def detect_and_track(self, frame, tracking_anchor_indices, tracking_anchor_weights, tracking_object_classes):
        with_tracking = len(tracking_anchor_indices) > 0
        if with_tracking:
            tracking_indices = mx.nd.array(tracking_anchor_indices, ctx=self._ctx)
            tracking_weights = mx.nd.array(tracking_anchor_weights, ctx=self._ctx)
            tracking_classes = mx.nd.array(tracking_object_classes.reshape((-1, 1)), ctx=self._ctx)
        else:
            tracking_classes = tracking_indices = tracking_weights = self._dummy_ti
        ids, scores, bboxes, voted_tracking_bboxes, detection_anchor_indices = self.detector.run_detection(frame, tracking_indices, tracking_weights, tracking_classes)
        valid_det_num = (scores > self._detector_thresh).sum().astype(int).asnumpy()[0]
        if valid_det_num > 0:
            valid_scores = scores[:valid_det_num]
            valid_bboxes = bboxes[:valid_det_num, :]
            valid_classes = ids[:valid_det_num, :]
            detection_output = mx.nd.concat(valid_bboxes, valid_scores, valid_classes, dim=1).asnumpy()
            anchor_indices_output = detection_anchor_indices[:valid_det_num, :]
        else:
            detection_output = np.array([])
            anchor_indices_output = np.array([])
        tracking_response = voted_tracking_bboxes.asnumpy() if with_tracking else np.array([])
        return detection_output, anchor_indices_output, tracking_response, {}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

这里引入的是COCO数据集的已经训练好的目标检测的模型

二、引入SMOT跟踪类

class SMOTTracker:
    def __init__(self,
                 motion_model='no',
                 anchor_array=None,
                 use_motion=True,
                 tracking_classes=[],
                 match_top_k=10,
                 track_keep_alive_thresh=0.1,
                 new_track_iou_thresh=0.3,
                 track_nms_thresh=0.5,
                 gpu_id=0,
                 anchor_assignment_method='iou',
                 joint_linking=False,
                 tracktor=None,classes=[]
                 ):
        self.use_motion = use_motion
        self.tracks = []
        self.all_track_id = 0
        self.pending_index = []
        self.conf_score_thresh = 0.1
        self.anchor_array = anchor_array
        self.next_frame_pred_index = []
        self.next_frame_pred_weights = []
        self.next_frame_pred_bbox = []
        self.waiting_update_tracks = []
        self.next_frame_ref_bbox = []
        self.last_frame = None
        self.k = match_top_k
        self.keep_alive = track_keep_alive_thresh
        self.new_track_iou_thresh = new_track_iou_thresh
        self.track_nms_thresh = track_nms_thresh
        self.frame_cache = None
        self.mx_ctx = mx.gpu(gpu_id)
        self.anchor_assignment_method = anchor_assignment_method
        self.joint_linking = joint_linking

        if len(tracking_classes) == 0:
            raise ValueError("Unknown tracking classes, please let us know what object you want to track")
        self.coco_class_set = set(classes)
        self.coco_class2index_dict = {i:name for i, name in enumerate(classes)}
        self.class_set = set(tracking_classes)
        for class_name in tracking_classes:
            if class_name not in self.coco_class_set:
                raise ValueError("Your cunstom class {} is not supported, only COCO classes are currently supported,\
                 the classes are {}".format(class_name, classes))

        if motion_model == 'farneback':
            self.motion_estimator = FarneBeckFlowMotionEstimator()
        elif motion_model == 'no':
            self.motion_estimator = DummyMotionEstimator()
        else:
            raise ValueError("Unknown motion model: {}".format(motion_model))

    def process_frame_sequence(self, frame_iterator, tracktor):
        for frame_id, frame in frame_iterator:
            print('Processing Frame ID: {}'.format(frame_id))
            t_iter_start = time.time()

            # STEP 0: Prepare the tracktor with the new frame data
            motion_pred_data = tracktor.prepare_for_frame(frame)

            # STEP 1: Predict the new locations of the tracked bounding boxes in the tracker
            tracking_anchor_indices, tracking_anchor_weights, tracking_classes = self.motion_predict(frame, motion_pred_data)

            # STEP 2: Run the tracktor
            detection_bboxes, detection_anchor_indices, tracking_bboxes, extra_info \
                = tracktor.detect_and_track(frame,
                                            tracking_anchor_indices,
                                            tracking_anchor_weights,
                                            tracking_classes
                                            )
            if self.anchor_array is None:
                self.set_anchor_array(tracktor.anchors())

            # STEP 3: Update the tracker with detector responses
            self.update(detection_bboxes,
                        tracking_bboxes,
                        detection_anchor_indices,
                        tracking_anchor_indices,
                        tracking_anchor_weights,
                        tracking_classes,
                        extra_info)

            # yield the results of this frame
            results = self._produce_frame_result()

            elapsed = time.time() - t_iter_start
            print("Total Tracking Runtime: {:2.4f} msec, {:.01f} FPS".format(elapsed * 1000, 1 / elapsed))
            yield frame_id, results

    @timeit
    def set_anchor_array(self, anchor_arracy):
        self.anchor_array = anchor_arracy
        self.anchor_tensor = mx.nd.array(self.anchor_array, ctx=self.mx_ctx, dtype=np.float32)

    @timeit
    def motion_predict(self, new_frame: np.ndarray, motion_pred_data):
        # STEP 1: Find all active tracks
        active_track_boxes = []
        active_track_indices = []
        active_track_anchor_indices = []
        active_track_anchor_weights = []
        active_track_classes = []

        t_active = time.time()
        for track_idx, track in enumerate(self.tracks):
            if track.is_active():
                active_track_boxes.append(track.mean)
                active_track_indices.append(track_idx)
                src_idx, src_weights = track.source
                active_track_anchor_indices.append(src_idx)
                active_track_anchor_weights.append(src_weights)
                active_track_classes.append([track.class_id])
                print("active track {} with age: {}".format(track.track_id, track.age))
        active_track_boxes = np.array(active_track_boxes)
        active_track_anchor_indices = np.array(active_track_anchor_indices)
        active_track_anchor_weights = np.array(active_track_anchor_weights)
        e_active = time.time() - t_active
        print('find active runtime: {:.05f}'.format(e_active))

        if len(active_track_boxes) > 0:
            # The following steps only happen if we have something to track

            # STEP 2: Warp the boxes according to flow
            predicted_track_boxes = self._motion_prediction(new_frame, active_track_boxes,
                                                            active_track_anchor_indices,
                                                            active_track_anchor_weights,
                                                            motion_pred_data,
                                                            skip=not self.use_motion)


            # STEP 3: Assign the warped boxes to anchor compositions
            tracking_anchor_indices, tracking_anchor_weights, tracking_anchor_validity = self._assign_box_to_anchors(
                predicted_track_boxes, method=self.anchor_assignment_method)

            # remove tracks becoming invalid after motion prediction
            invalid_track_numbers = np.nonzero(1 - tracking_anchor_validity)[0]
            print("{}/{} tracks become invalid after motion prediction".format(len(invalid_track_numbers), len(active_track_boxes)))

            for i_invalid in invalid_track_numbers:
                self.tracks[active_track_indices[i_invalid]].mark_missed()

            # keep the valid tracks for re-detection
            valid_track_numbers = np.nonzero(tracking_anchor_validity)[0]
            self.next_frame_pred_index = tracking_anchor_indices[valid_track_numbers, ...]
            self.next_frame_pred_weights = tracking_anchor_weights[valid_track_numbers, ...]
            self.next_frame_pred_bbox = predicted_track_boxes[valid_track_numbers, ...]
            self.next_frame_pred_class = np.array(active_track_classes)[valid_track_numbers, ...]

            active_track_indices = np.array(active_track_indices)[valid_track_numbers, ...]
            active_track_boxes = active_track_boxes[valid_track_numbers, ...]

        else:
            # skip flow computation if there is no active track
            # just save the frame in cache
            predicted_track_boxes = self._motion_prediction(new_frame,
                                                            active_track_boxes,
                                                            active_track_anchor_indices, active_track_anchor_weights,
                                                            motion_pred_data, skip=True)

            assert len(predicted_track_boxes) == 0

            self.next_frame_pred_index = np.array([])
            self.next_frame_pred_weights = np.array([])
            self.next_frame_pred_bbox = np.array([])
            self.next_frame_pred_class = np.array([])

        self.waiting_update_tracks = active_track_indices
        self.next_frame_ref_bbox = active_track_boxes

        return self.next_frame_pred_index, self.next_frame_pred_weights, self.next_frame_pred_class

    @timeit
    def update(self, new_detections: np.ndarray, tracking_predictions: np.ndarray,
               detection_anchor_indices: np.ndarray,
               tracking_anchor_indices: np.ndarray, tracking_anchor_weights: np.ndarray,
               tracking_classes: np.ndarray,
               extra_info: dict = None):
        # pylint: disable=too-many-nested-blocks
        t_pose_processing = time.time()

        print("tracking predictions 's shape is {}".format(tracking_predictions.shape))
        print(tracking_predictions)
        print(self.waiting_update_tracks)

        detection_landmarks = extra_info['detection_landmarks'] if 'detection_landmarks' in extra_info else None
        tracking_landmarks = extra_info['tracking_landmarks'] if 'tracking_landmarks' in extra_info else None

        for t in self.tracks:
            t.predict()

        # STEP 1: track level NMS
        still_active_track_pred_indices = []
        still_active_track_indices = []

        if len(tracking_predictions) > 0:

            # class wise NMS
            keep_set = set()
            for c in set(tracking_classes.ravel().tolist()):
                class_pick = np.nonzero(tracking_classes == c)[0]
                keep_tracking_pred_nms_indices = nms_fallback(tracking_predictions[class_pick, ...], self.track_nms_thresh)
                for i_keep in keep_tracking_pred_nms_indices:
                    keep_set.add(class_pick[i_keep])

            still_active_track_pred_indices = []
            for i_pred, i_track in enumerate(self.waiting_update_tracks):
                if i_pred in keep_set:
                    self.tracks[i_track].update(tracking_predictions[i_pred, :],
                                                (tracking_anchor_indices[i_pred, :], tracking_anchor_weights[i_pred, :]),
                                                tracking_landmarks[i_pred, :] if tracking_landmarks is not None else None)
                else:
                    # suppressed tracks in the track NMS process will be marked as Missing
                    self.tracks[i_track].mark_missed()

                if self.tracks[i_track].is_active():
                    still_active_track_pred_indices.append(i_pred)
                    still_active_track_indices.append(i_track)

        # STEP 2: Remove New Detection Overlapping with Tracks
        if len(still_active_track_pred_indices) > 0 and len(new_detections) > 0:
            active_tracking_predictions = tracking_predictions[still_active_track_pred_indices, :]
            det_track_max_iou = bbox_iou(new_detections[:, :4], active_tracking_predictions[:, :4])
            same_class = new_detections[:, -1:] == (tracking_classes[still_active_track_pred_indices, :].T)
            # suppress all new detections that have high IOU with active tracks
            affinity = (det_track_max_iou * same_class).max(axis=1)
            keep_detection_indices = np.nonzero(affinity <= self.new_track_iou_thresh)[0]
        else:
            # otherwise simply keep all detections
            keep_detection_indices = list(range(len(new_detections)))
            active_tracking_predictions = np.array([])

        # STEP 3: New Track Initialization
        if len(keep_detection_indices) > 0:

            active_new_detections = new_detections[keep_detection_indices, :]
            # (Optional) STEP 3.a: Perform joint linking of body and head
            if self.joint_linking:
                tracking_classes = np.array(tracking_classes)
                body2face_link, face2body_link = \
                    self._link_face_body(active_new_detections,
                                         extra_info['detection_keypoints'][keep_detection_indices],
                                         active_tracking_predictions,
                                         extra_info['tracking_keypoints'][still_active_track_pred_indices],
                                         tracking_classes[still_active_track_pred_indices]
                                         )
            else:
                body2face_link, face2body_link = None, None

            new_tracks = []
            for idx, i_new_track in enumerate(keep_detection_indices):
                new_track = Track(new_detections[i_new_track, :4], self.all_track_id,
                                  (detection_anchor_indices[i_new_track, :], np.array([1])),
                                  keep_alive_thresh=self.keep_alive, class_id=new_detections[i_new_track, -1],
                                  attributes=detection_landmarks[i_new_track, :] if detection_landmarks is not None else None)
                if self.joint_linking:
                    if new_track.class_id == 0:
                        # new face track
                        if idx in face2body_link[0]:
                            print(idx, i_new_track, '0')
                            body_idx = face2body_link[0][idx]
                            if idx > body_idx:
                                new_track.link_to(new_tracks[body_idx])
                        elif idx in face2body_link[2]:
                            print(idx, i_new_track, '1')
                            body_idx = face2body_link[2][idx]
                            new_track.link_to(self.tracks[still_active_track_indices[body_idx]])

                    if new_track.class_id == 1:
                        # new body track
                        if idx in body2face_link[0]:
                            face_idx = body2face_link[0][idx]
                            if idx > face_idx:
                                new_track.link_to(new_tracks[face_idx])
                        elif idx in body2face_link[2]:
                            face_idx = body2face_link[2][idx]
                            new_track.link_to(self.tracks[still_active_track_indices[face_idx]])

                self.all_track_id += 1
                self.tracks.append(new_track)
                new_tracks.append(new_track)

        elapsed_post_processing = time.time() - t_pose_processing
        print("total tracklets to now is {}, post-processing time: {:.05f} sec".format(
            self.all_track_id, elapsed_post_processing))

    @property
    def active_tracks(self):
        for t in self.tracks:
            if t.is_active():
                yield t

    def _motion_prediction(self, new_frame, tracked_boxes,
                           tracked_boxes_anchor_indices, tracked_boxes_anchor_weights,
                           motion_pred_data, skip=False):
        if self.frame_cache is None:
            # this is the first frame
            self.frame_cache = self.motion_estimator.initialize(new_frame, motion_pred_data)
            predicted_bboxes = tracked_boxes
        else:
            # this is not the first frame
            predicted_bboxes, self.frame_cache = self.motion_estimator.predict_new_locations(
                self.frame_cache, tracked_boxes, new_frame, motion_pred_data,
                tracked_boxes_anchor_indices=tracked_boxes_anchor_indices,
                tracked_boxes_anchor_weights=tracked_boxes_anchor_weights,
                skip=skip)

        return predicted_bboxes

    def _assign_box_to_anchors(self, boxes: np.ndarray, method: str = 'avg', min_anchor_iou: float = 0.1):
        t_start = time.time()

        t_iou = time.time()
        gpu_boxes = mx.nd.array(boxes, ctx=self.mx_ctx)
        anchor_track_iou = gpu_iou(self.anchor_tensor, gpu_boxes)
        elapsed_iou = time.time() - t_iou
        print("iou computation runtime: {:.05f}".format(elapsed_iou))

        # get the top-k closest anchors instead of 1
        if method == 'max':
            tracking_anchor_ious, tracking_anchor_indices = mx.nd.topk(anchor_track_iou, axis=0, k=1,
                                                                       ret_typ='both', dtype='int32')
            tracking_anchor_ious = tracking_anchor_ious.T.asnumpy()
            tracking_anchor_indices = tracking_anchor_indices.T.asnumpy()
            tracking_anchor_weights = np.ones_like(tracking_anchor_indices)
        elif method == 'avg':
            tracking_anchor_ious, tracking_anchor_indices = mx.nd.topk(anchor_track_iou, axis=0, k=self.k,
                                                                       ret_typ='both', dtype='int32')
            tracking_anchor_ious = tracking_anchor_ious.T.asnumpy()
            tracking_anchor_indices = tracking_anchor_indices.T.asnumpy()
            tracking_anchor_weights = np.ones_like(tracking_anchor_indices) / self.k
        elif method == 'iou':
            t_sort = time.time()

            tracking_anchor_ious, tracking_anchor_indices = mx.nd.topk(anchor_track_iou, axis=0, k=self.k,
                                                                       ret_typ='both', dtype='int32')
            tracking_anchor_ious = tracking_anchor_ious.T.asnumpy()
            tracking_anchor_indices = tracking_anchor_indices.T.asnumpy()

            e_sort = time.time() - t_sort
            print('sorting time: {:.05f}'.format(e_sort))
            tracking_anchor_weights = tracking_anchor_ious / tracking_anchor_ious.sum(axis=1)[..., None]
        else:
            raise ValueError("unknown anchor assignment method")

        max_track_anchor_ious = tracking_anchor_ious.max(axis=1)
        tracking_anchor_validity = max_track_anchor_ious >= min_anchor_iou

        elapsed_assign = time.time() - t_start
        print("assigment runtime: {}".format(elapsed_assign))
        return tracking_anchor_indices, tracking_anchor_weights, tracking_anchor_validity

    @timeit
    def _produce_frame_result(self):
        tracked_objects = []
        for track in self.active_tracks:
            box = {
                'left': track.mean[0],
                'top': track.mean[1],
                'width': track.mean[2] - track.mean[0],
                'height': track.mean[3] - track.mean[1]
            }
            tid = track.display_id
            age = track.age
            classId = track.class_id
            obj = {
                'bbox': box,
                'track_id': tid,
                'age': age,
                'class_id':classId,
                'class_name':self.coco_class2index_dict[classId]
            }
            if track.attributes is not None:
                obj['landmarks'] = track.attributes
            if obj["class_name"] in self.class_set:
                tracked_objects.append(obj)
        return tracked_objects

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379

三、调用方式

tracktor = GluonSSDMultiClassTracktor(gpu_id=0, detector_thresh=0.2, model_name='ssd_512_vgg16_atrous_coco', use_pretrained=True, param_path="", data_shape=512,classes = classes_names)
tracker = SMOTTracker(match_top_k=5,
                                  motion_model='farneback',
                                  use_motion=True,
                                  tracking_classes=["person"],
                                  anchor_assignment_method='iou',
                                  track_keep_alive_thresh=0.3,classes=classes_names)
1
2
3
4
5
6
7

效果图：

同一颜色代表同一目标