DyGLIP: A Dynamic Graph Model with Link Prediction for Accurate Multi-Camera Multiple Object Tracking

Abstract

Multi-Camera Multiple Object Tracking (MC-MOT) is a significant computer vision problem due to its emerging applicability in several real-world applications. Despite a large number of existing works, solving the data association problem in any MC-MOT pipeline is arguably one of the most challenging tasks. Developing a robust MC-MOT system, however, is still highly challenging due to many practical issues such as inconsistent lighting conditions, varying object movement patterns, or the trajectory occlusions of the objects between the cameras. To address these problems, this work, therefore, proposes a new Dynamic GraphModel with Link Prediction (DyGLIP) approach 1 to solve the data association task. Compared to existing methods, our new model offers several advantages, including better feature representations and the ability to recover from lost tracks during camera transitions. Moreover, our model works gracefully regardless of the overlapping ratios between the cameras. Experimental results show that we outperform existing MC-MOT algorithms by a large margin on several practical datasets. Notably, our model works favorably on online settings but can be extended to an incremental approach for large-scale datasets.

LMGP: Lifted Multicut Meets Geometry Projections for Multi-Camera Multi-Object Tracking

“…Single-Camera Tracklets Centralized Representation Online LMGP (Ours) MLMRF [27] STVH [41] DMCT [46] GLMB [31] DyGLIP [32] Table 5. Comparison of our LMGP w.r.t.…”

“…Recent approaches include a semi-online Multi-Label Markov Random Field (MLMRF) method [27], where the ensuing optimization problem over single detections is solved through alpha-expansion [4] and a non-negative matrix factorization approach (TRACTA) for grouping tracklets across cameras [16]. In another direction, DyGLIP [32] formulates the data association problem for multi-camera as link prediction on a graph whose nodes are tracklets. While these methods have demonstrated promising performance in some datasets, they are affected by ID-switch errors in the tracklet proposal generation, especially in cluttered or crowded scenes such as [7].…”

“…Baselines: For each dataset, we compare against the most recent approaches for which we found experimental results. In particular, we compare against KPS-DO, KSP-DO-ptrack [7], GLMB-YOLOv3, GLMB-DO [31], DMCT, DMCT Stack [46], HJMV [20], STVH [41], MLMRF [27], HCT [43], STP [44], TRACTA [16] and DyGLIP [32]. All baseline performances are taken from the respective original papers.…”

“…Leveraging this property, recent papers have pursued two principal approaches: single viewbased and centralized representation methods. In the first strategy, [16,27,32,44], a two-step procedure is followed: 1) generating local tracklets of all the targets within each camera; 2) matching local tracklets that belong to the same target across cameras through computing affinity costs and using a global optimization framework. While this framework brings benefits via a reduced hypothesis space and allows to design motion-based features, its main drawbacks lie in ID-Switch errors contained inside local tracklets, i.e., detections of distinct objects are grouped into the same trajectory (Figure 3a).…”

OTAdapt: Optimal Transport-based Approach For Unsupervised Domain Adaptation

“…Deep learning-based image recognition studies have been recently achieving very accurate performance in visual applications, e.g. image classification [1], [2], [3], face recognition, [4], [5], [6], [7], [8], image synthesis [9], [10], [11], [12], [13], [14], action recognition [15], [16], semantic segmentation [17], [18]. However, these methods assume the testing images from the same distribution as the training images, therefore, these deep learning-based models are likely to fail when performing in real data in the new domains.…”