GTATrack: Multi-Domain Tracking System

• GTATrack is a multi-domain tracking system that integrates advanced computer vision and GNSS analytics to monitor sports, vehicles, railways, and particle beams.
• It employs a two-stage pipeline using Deep Expansion IoU for local association and Global Tracklet Association for offline trajectory refinement to enhance tracking accuracy.
• The system features plug-and-play integration with various detectors and includes a pseudo-labeling strategy to improve detection recall under challenging conditions.

GTATrack refers to tracking systems and technologies spanning multiple domains: advanced multi-object tracking in sports video (notably with Deep Expansion IoU and Global Tracklet Association), vehicle fleet management using GNSS/GSM hardware plus web-based analytics, track geometry measurement in railways, and high-rate beam tracking for high-energy physics. The term “GTATrack” and related systems have achieved prominent results across soccer analytics, transportation logistics, railway monitoring, and particle tracking, each leveraging domain-specific algorithms, architectures, and instrumentation.

1. Multi-Object Tracking in Sports: GTATrack Framework

The GTATrack system, as developed for visual multi-object tracking (MOT) in soccer, is a two-stage, hierarchical tracking pipeline integrating Deep Expansion IoU (Deep-EIoU) and a global refinement module (GTA-Link) (Jian et al., 31 Jan 2026, Sun et al., 2024). The approach is designed for the challenging video domain characterized by static, fisheye cameras, severe geometric distortion, frequent occlusion, and uniform player appearance.

The processing pipeline operates as follows:

1. Detection: Input frames are processed by a detector (e.g., YOLOv11x or YOLOX), yielding bounding boxes and scores for candidate targets.
2. Feature Extraction: Each detection is encoded via a re-identification (ReID) network (e.g., OSNet), producing per-box appearance descriptors.
3. Local Association (Deep-EIoU): Frame-to-frame matching is performed with the Deep-EIoU matcher, which replaces standard IoU by stacking box-expanded IoUs and applying an MLP over the resulting vector, fused with appearance similarity, allowing robust tracklet construction without explicit motion prediction.
4. Global Tracklet Association (GTA-Link): The initial set of short tracklets output by Deep-EIoU is refined offline through GTA-Link, a clustering-based mechanism that splits mixed-identity tracklets and merges fragmented trajectories based on global appearance/spatio-temporal affinity.

This architecture, distinct in its separation of local and global reasoning, achieves high HOTA scores and low false positive/ID switch rates under difficult tracking conditions (Jian et al., 31 Jan 2026).

2. Deep Expansion IoU and Online Tracklet Formation

Deep-EIoU provides motion-agnostic data association by leveraging not only traditional IoU but a vector of IoUs computed after iterative expansions of bounding box coordinates. Formally, for bounding boxes $b_1$, $b_2$, and expansion factors $r_k$, compute $$EIoU_k(b_1, b_2) = \text{IoU}(\text{Expand}(b_1, r_k), \text{Expand}(b_2, r_k))$$ for $k=1,…,K$. The resulting vector and deep appearance similarity $(f_1 \cdot f_2)$ are aggregated by a learned scoring function and thresholded for association.

This online phase solves a linear assignment problem per frame (Hungarian algorithm) to maintain short, reliable tracklets, robust against erratic motion and non-linear displacements typical in soccer datasets (Jian et al., 31 Jan 2026).

3. Global Tracklet Association (GTA-Link): Offline Trajectory Refinement

The second stage, GTA-Link, operates at the trajectory level, addressing identity switches and tracking fragmentation. It treats short tracklets as graph nodes, with pairwise dissimilarities given by a convex combination of average appearance distance, temporal gap, and Deep-EIoU overlap in a bridge window. A standard cost function is

$$w_{ij} = \alpha D_\mathrm{app} + \beta \frac{\Delta t}{T_\mathrm{max}} + \gamma (1 - \text{Deep-EIoU}_\text{bridge})$$

with constraints ensuring temporal non-overlap and spatial plausibility, and associations solved by ILP or agglomerative clustering.

The splitting of impure tracklets is performed by a modified DBSCAN algorithm over embedding sequences with fixed hyperparameters $(s, \epsilon, k)$. Fragments are greedily merged globally with appearance- and context-based constraints until no further mergeable pairs exist (distance exceeding $\alpha$ threshold) (Sun et al., 2024).

4. Integration as a Plug-and-Play Refinement Module

GTATrack is tracker- and detector-agnostic, accepting any tracker’s output (SORT, ByteTrack, Deep-EIoU, etc.) and associated per-detection ReID features. The module operates entirely as post-processing, requiring no modification or retraining of upstream components. The pipeline is compatible with both learned and oracle (ground-truth) detections, and can be integrated into evaluation toolchains as a refinement/augmentation stage, providing ID-reindexed, globally consistent tracks (Sun et al., 2024).

5. Pseudo-Labeling Strategy for Small and Distorted Targets

To address poor detection recall on small, occluded, or highly distorted targets, GTATrack incorporates a pseudo-labeling bootstrapping loop. Initial detectors are run on unlabeled data to harvest confident detections ($\tau_p \geq 0.7$), which are re-used as “pseudo ground-truth” for additional supervised training. The overall loss function penalizes both true and pseudo-labeled objectness predictions, weighted by $b_2$0 to control their impact. This semi-supervised approach significantly reduces false negatives and improves overall recall, an effect quantitatively confirmed in ablation studies (Jian et al., 31 Jan 2026).

6. Evaluation Metrics, Empirical Performance, and Implementation

Performance is evaluated via HOTA (Higher Order Tracking Accuracy), which balances detection, localization, and association accuracy. In publicly benchmarked contexts:

• SoccerTrack 2025: HOTA = 0.60 (1st place), FP = 982, IDSW = 331.5, with localization accuracy 0.84 and association accuracy 0.47.
• SportsMOT and SoccerNet 2023: HOTA improvements over baselines $b_2$1 3.5 points, with association accuracy (AssA) and IDF1 scores rising $b_2$2–$b_2$3 points and switch counts dropping by $b_2$41,900 (Sun et al., 2024, Jian et al., 31 Jan 2026).

Key implementation details include detector training (YOLOv11x or YOLOX), OSNet ReID for embeddings, Deep-EIoU expansion $b_2$5, clustering (DBSCAN or ILP), single-threaded Python, and mixed-precision–optimized high-resolution image processing. Open-source codebases are provided for reproducibility.

7. Comparative Context: GTATrack Across Other Domains

The term “GTATrack” also refers to systems outside the sports video tracking context:

• Fleet Management: GTATrack denotes the Rad100 GPS/GLONASS hardware integrated with PayaRadyab web software, providing vehicle tracking, analytics (fuel, speed, route efficiency), and global coverage. The system uses custom hardware (ARM Cortex-M3, SIM908) and robust, secure client/server communication for high-uptime, scalable deployment (Saghaei, 2016).
• Track Geometry Measurement: GTMS systems, implementing computer vision and inertial navigation for railway track monitoring, employ methodologies that mirror GTATrack’s principles at the sensor fusion and data pipeline level, though targeting geometric parameter estimation (gauge, alignment, twist) rather than object tracking (Escalona, 2020).
• Beam Tracking in Particle Physics: The NA62 GigaTracKer delivers sub-100 ps time resolution at MHz/mm² flux using hybrid pixel ASICs, cooling, and bandwidth-optimized electronics for 4D beam tracking, addressing analogous challenges of association and time/space resolution in the high-rate regime (Rinella et al., 2019).

While the core GTATrack method as established in (Sun et al., 2024, Jian et al., 31 Jan 2026) is rooted in computer vision for sports analytics, the consistent focus on trajectory association, identity continuity, and plug-and-play refinement is a recurring thread through these diverse applications.