GRAP-MOT: Unsupervised Graph-based Position Weighted Person Multi-camera Multi-object Tracking in a Highly Congested Space

TL;DR

GRAP-MOT tackles multi-camera MOT in highly congested closed spaces by online graph-based identity updates that fuse head-based features and, when available, position estimates. The method builds a cross-camera tracklet graph with edge similarities $s_{ij}^f$, $s_{ij}^p$, and edge occurrences $o_{ij}$, updated through $I_{ij}$ and refined via graph partitioning with $Q$ and repulsion rules. Through extensive internal and CAMPUS dataset experiments, the authors identify the best single-camera tracker (SORT), best head feature (ResNet50), and a greedy modularity based partitioning scheme, while showing the crucial role of position estimation in achieving strong cross-camera identity consistency as measured by IDF1. They also argue that IDF1 provides a more faithful MOT quality measure than MOTA in congested scenarios, discuss limitations, and provide public access to code and data for reproducibility and benchmarking.

Abstract

GRAP-MOT is a new approach for solving the person MOT problem dedicated to videos of closed areas with overlapping multi-camera views, where person occlusion frequently occurs. Our novel graph-weighted solution updates a person's identification label online based on tracks and the person's characteristic features. To find the best solution, we deeply investigated all elements of the MOT process, including feature extraction, tracking, and community search. Furthermore, GRAP-MOT is equipped with a person's position estimation module, which gives additional key information to the MOT method, ensuring better results than methods without position data. We tested GRAP-MOT on recordings acquired in a closed-area model and on publicly available real datasets that fulfil the requirement of a highly congested space, showing the superiority of our proposition. Finally, we analyzed existing metrics used to compare MOT algorithms and concluded that IDF1 is more adequate than MOTA in such comparisons. We made our code, along with the acquired dataset, publicly available.

Figures (5)

• Figure 1: Iconographic visualization of camera distribution on the closed-area model frame. The exemplary person's multi-object tracking on different camera views is included in different colors, where left/right (a and c) are typical views and (b) in the middle is a fish-eye type view.
• Figure 2: Overview of the GRAP-MOT Multi-Camera Multi-Object Tracking Pipeline. Video frames from multiple cameras are processed sequentially. A head detection module outputs bounding box coordinates for each detected individual. Within each camera view, a tracking algorithm assigns temporary identity labels to detections, forming short-term trajectories (tracklets). A position estimation module projects these tracklets onto a common spatial grid. All tracklets are then represented as nodes in a graph, where edges connect tracklets originating from different cameras. An importance coefficient is computed for each edge to quantify the likelihood of cross-camera identity correspondence. Tracklets are clustered across cameras into groups, with group sizes limited by the total number of cameras. Finally, unified identity labels are assigned within each group, yielding consistent tracking across all camera views.
• Figure 3: Overview of graph update, partitioning and repulsion processes.
• Figure 4: Example frame from the Garden2 recording illustrating discrepancies between IDF1 and MOTA. (a) GRAP-MOT final pipeline with a few ID mismatches, yielding IDF1 = 38.66 and MOTA = 48.09. (b) ReST with spatial and temporal graphs trained on Garden2 shows low IDF1 = 32.2 but high MOTA = 85.5 due to duplicated IDs. Color coding: green – correctly tracked detections across cameras; orange – mismatch on one camera; red – mismatch on two cameras; blue – repeated ID on one camera; grey – detections missing in other frames.
• Figure 5: Example frame from the internal gt6_task10 recording highlighting the mismatch between IDF1 and MOTA. a) GRAP-MOT final pipeline, where both IDF1=97.607 and MOTA=95.214 indicate relatively correct results, b) GRAP-MOT testing GNI community detection algorithm for which mismatches are visible in IDF1=39.567 while MOTA=86.463 is still very high. Green: all detections tracked correctly across the multi-camera views, orange: MOT mismatch on one camera, red: MOT mismatch on two cameras and grey: detections not tracked on other frames.