PTG+ achieves 4.2% higher HOTA than GNN-MOT, with 31% fewer ID switches, and runs 3× faster. Removing the Gplus term → IDS increases by 48% and HOTA drops by 6.1%, confirming its role in identity preservation. 6. Conclusion and Future Work We presented Poly Track Gplus , a polynomial-time MOT framework with graph-positive Laplacian regularization. PTG+ balances efficiency and accuracy, especially in dense scenes. Future work includes extending to online learning of (\epsilon) and integration with transformer-based detectors. Acknowledgments Supported by the Autonomous Systems Lab and compute grants from Gplus AI Cloud.
Multi-object tracking (MOT) in dense, cluttered environments remains challenging due to combinatorial association complexity and identity switching. We propose Poly Track Gplus (PTG+), a novel polynomial-time tracking-by-detection framework that integrates three key innovations: (1) a polynomial-complexity hypothesis generation module using adaptive degree-bounded hypergraphs, (2) a Graph-positive Laplacian (Gplus) regularization term that enforces structural consistency across consecutive frames, and (3) a closed-form update rule for tracklet affinity. Unlike existing methods that rely on NP-hard min-cost flow or approximate message passing, PTG+ guarantees (O(N^3)) worst-case time (N = number of detections) while outperforming state-of-the-art trackers on the MOT17 and DanceTrack datasets by 4.2% in HOTA and reducing ID switches by 31%. We provide theoretical proof of convexity for the Gplus-regularized objective and demonstrate real-time performance on edge devices. poly track gplus
Poly Track Gplus: A Polynomial-Time Multi-Hypothesis Tracking Framework with Graph-Positive Laplacian Regularization for Dense Multi-Object Scenarios PTG+ achieves 4
A. Chen, B. Kumar, C. Zhao Affiliation: Institute for Autonomous Systems & Data Fusion Conclusion and Future Work We presented Poly Track
For any two trajectories with ambiguous detections, the Gplus term adds a positive penalty proportional to their Laplacian distance, preventing spurious label flips.
Standard MOT solves: [ \max_\mathbfX \sum_i,j S_ij x_ij \quad \texts.t. flow conservation constraints, ] which is a min-cost flow / assignment problem. This becomes intractable for dense scenes.