Multi-target Tracking in Multiple Non-overlapping Cameras Using Fast-Constrained Dominant Sets

Abstract

In this paper, a unified three-layer hierarchical approach for solving tracking problems in multiple non-overlapping cameras is proposed. Given a video and a set of detections (obtained by any person detector), we first solve within-camera tracking employing the first two layers of our framework and, then, in the third layer, we solve across-camera tracking by merging tracks of the same person in all cameras in a simultaneous fashion. To best serve our purpose, a constrained dominant sets clustering (CDSC) technique, a parametrized version of standard quadratic optimization, is employed to solve both tracking tasks. The tracking problem is caste as finding constrained dominant sets from a graph. That is, given a constraint set and a graph, CDSC generates cluster (or clique), which forms a compact and coherent set that contains a subset of the constraint set. The approach is based on a parametrized family of quadratic programs that generalizes the standard quadratic optimization problem. In addition to having a unified framework that simultaneously solves within-and across-camera tracking, the third layer helps link broken tracks of the same person occurring during within-camera tracking. A standard algorithm to extract constrained dominant set from a graph is given by the so-called replicator dynamics whose computational complexity is quadratic per step which makes it handicapped for large-scale applications. In this work, we propose a fast algorithm, based on dynamics from evolutionary game theory, which is efficient and salable to large-scale real-world applications. We have tested this approach on a very large and challenging dataset (namely, MOTchallenge DukeMTMC) and show that the proposed framework outperforms the current state of the art. Even though the main focus of this paper is on multi-target tracking in non-overlapping cameras, proposed approach can also be applied to solve re-identification problem. Towards that end, we also have performed experiments on MARS, one of the largest and challenging video-based person re-identification dataset, and have obtained excellent results. These experiments demonstrate the general applicability of the proposed framework for non-overlapping across-camera tracking and person re-identification tasks.

Eliminating Exposure Bias and Metric Mismatch in Multiple Object Tracking

“…Remarkably, on DukeMTMC dataset, even though we ignored appearance for the purpose of this comparison, our approach also outperforms or rivals some the methods that exploit it [46,49]. This shows that our training procedure is powerful enough to overcome this serious handicap.…”

“…Most state-of-the-art approaches that use sequence models rely on one of two optimization techniques, hierarchical clustering for data association [49,59,46,34,18,25] or multiple hypothesis tracking [56,27,10]. The former involves valid groups of observations without shared hypotheses while the latter allows for conflicting sets of hypotheses to be present until the final solution is found.…”

“…• CDSC [49] uses domination set clustering to perform within-and across-camera tracking. It employs image features from ResNet-50 [16] pre-trained on ImageNet.…”

Features for Multi-target Multi-camera Tracking and Re-identification

“…Overall results are presented in Tables 1 and 3. Our method DeepCC improves the multi-camera IDF1 accuracy w.r.t to the previous state of the art MTMC CDSC [68] by 22 and 17.6 points for the test-easy and test-hard sequences, respectively. For the single-camera easy and hard sequences, the IDF1 improvement is 12.2 and 13.5 points, and MOTA improves by 16.6 and 10.4 points.…”

“…We nonetheless outperform all methods on IDF1, IDR and MOTA. BIPCC [57] PT BIPCC [49] MTMC CDSC [68] MYTRACKER [72] MTMC ReID [79] † DeepCC BIPCC [57] PT BIPCC [49] MTMC CDSC [68] MYTRACKER [72] MTMC ReID [79] † DeepCC BIPCC [57] PT BIPCC [49] MTMC CDSC [68] MYTRACKER [72] MTMC ReID [79] † DeepCC BIPCC [57] PT BIPCC [49] MTMC CDSC [68] MYTRACKER [72] MTMC ReID [ It is worth noting that our method achieves the highest identity recall IDR on all scenarios, and on nearly all single-camera sequences. Identity recall is Achille's heel for modern multi-target trackers, as they commonly fail to re-identify targets after occlusions [44].…”