Method step

models/tracker.py:53–136

(self, output_results)

    
    def step(self, output_results):
        scores = output_results["scores"]
        classes = output_results["labels"]
        bboxes = output_results["boxes"]  # x1y1x2y2
        track_bboxes = output_results["track_boxes"] if "track_boxes" in output_results else None # x1y1x2y2
        
        results = list()
        results_dict = dict()

        tracks = list()
        
        for idx in range(scores.shape[0]):
            if idx in self.tracks_dict and track_bboxes is not None:
                self.tracks_dict[idx]["bbox"] = track_bboxes[idx, :].cpu().numpy().tolist()

            if scores[idx] >= self.score_thresh:
                obj = dict()
                obj["score"] = float(scores[idx])
                obj["bbox"] = bboxes[idx, :].cpu().numpy().tolist()               
                results.append(obj)        
                results_dict[idx] = obj
        
        tracks = [v for v in self.tracks_dict.values()] + self.unmatched_tracks
        N = len(results)
        M = len(tracks)
        
        ret = list()
        unmatched_tracks = [t for t in range(M)]
        unmatched_dets = [d for d in range(N)]
        if N > 0 and M > 0:
            det_box   = torch.stack([torch.tensor(obj['bbox']) for obj in results], dim=0) # N x 4        
            track_box = torch.stack([torch.tensor(obj['bbox']) for obj in tracks], dim=0) # M x 4                
            cost_bbox = 1.0 - box_ops.generalized_box_iou(det_box, track_box) # N x M

            matched_indices = linear_sum_assignment(cost_bbox)
            unmatched_dets = [d for d in range(N) if not (d in matched_indices[0])]
            unmatched_tracks = [d for d in range(M) if not (d in matched_indices[1])]

            matches = [[],[]]
            for (m0, m1) in zip(matched_indices[0], matched_indices[1]):
                if cost_bbox[m0, m1] > 1.2:
                    unmatched_dets.append(m0)
                    unmatched_tracks.append(m1)
                else:
                    matches[0].append(m0)
                    matches[1].append(m1)

            for (m0, m1) in zip(matches[0], matches[1]):
                track = results[m0]
                track['tracking_id'] = tracks[m1]['tracking_id']
                track['age'] = 1
                track['active'] = 1
                pre_box = tracks[m1]['bbox']
                cur_box = track['bbox']
    #             pre_cx, pre_cy = (pre_box[0] + pre_box[2]) / 2, (pre_box[1] + pre_box[3]) / 2
    #             cur_cx, cur_cy = (cur_box[0] + cur_box[2]) / 2, (cur_box[1] + cur_box[3]) / 2
    #             track['vxvy'] = [cur_cx - pre_cx, cur_cy - pre_cy]
                ret.append(track)