Artificial Vision Project Work 2025/2026 University of Salerno - Dept. of Information Engineering, Electrical Engineering and Applied Mathematics

Authors: Antonio Apicella, Antonio Graziosi

Group: 16

This project presents a complete Computer Vision pipeline developed for the SoccerNet Video Understanding Benchmark Suite. The goal is to address two main tasks on real soccer match video clips:

1. Player Detection & Tracking: Uniquely detect and track players and referees on the field ("persons of interest"), maintaining consistent identity despite occlusions, rapid camera movements, and the absence of distinctive visual features (identical jerseys).
2. Behavior Analysis: Estimate player density in specific Regions of Interest (ROI) of the field for each frame.

The proposed solution is based on an optimized Tracking-by-Detection approach, which combines a state-of-the-art detector (YOLOv11x) configured for high sensitivity (High Recall), an adaptive Field Masking module for removing false positives on the sidelines, and a motion-based tracker (BoT-SORT with global camera compensation GMC), devoid of visual Re-Identification modules.

The system was evaluated on the Challenge Test Set, achieving an [email protected] of 0.742.

• High-Recall Detector: Use of YOLOv11x with 1088px input resolution and minimal confidence thresholds (0.1/0.2) to detect small objects and players blurred by motion.
• Adaptive Field Masking: Semantic segmentation algorithm combining HSV and LAB color spaces to dynamically identify the pitch and filter out stewards, photographers, and the audience, drastically reducing False Positives.
• "Pure Motion" Tracking: Implementation of BoT-SORT with Sparse Optical Flow (GMC) for camera motion compensation. The system does not use visual Re-ID, relying on robust geometric logic (match_thresh: 0.9) to avoid identity swaps between teammates.
• Behavior Analysis: Presence estimation in ROIs based on the geometric projection of the players' "feet point" (ground contact point).

├── configs/              # YAML and JSON configuration files (tracker, ROI, main)
├── models/               # Model weights (YOLO, ReID) and conversion scripts
├── output/               # Output folder for logs, videos, and results
├── SIMULATOR/            # Code for simulation and local testing
├── src/                  # Main source code
│   ├── behaviour/        # Module for behavior analysis
│   ├── evaluation/       # Evaluation module (HOTA, nMAE) using TrackEval
│   ├── tracker/          # Tracker logic (YOLO + BoT-SORT integration)
│   ├── utils/            # Utilities (Field Masking, BBox operations, visualization)
│   ├── visualizer/       # Video generation with overlay
│   └── main.py           # Pipeline entry point
└── tracking/             # Dataset (train, test, challenge)

1. Clone the repository:

git clone https://github.com/apiantonio/SoccerNet-tracking_AV2025-26
cd SoccerNet-tracking_AV2025-26

2. Install dependencies: It is recommended to use a virtual environment (venv or conda).

pip install -r requirements.txt

Main dependencies: ultralytics, opencv-python, numpy, pyyaml, scikit-image, pandas, trackeval.

3. Data Setup: Place the SoccerNet dataset in the tracking/ folder. The expected structure for each sequence is:

tracking/test/SNMOT-XXX/
├── img1/          # JPEG Frames
├── gt/gt.txt      # Ground Truth (optional for inference)
└── gameinfo.ini   # Metadata

The pipeline can be executed via the src/main.py script. You can specify which steps to perform (tracking, analysis, visualization, evaluation).

python src/main.py --config configs/main_config.yaml --step all --seq all

• --step: Choose between tracking, behaviour, eval, visualizer or all.
• --seq: Specify one or more sequences (e.g., SNMOT-116) or all to process the entire folder.
• --debug: Enables on-screen visualization (show_tracks, show_mask, show_behaviour).

Tracker parameters can be modified in the configs/botsort_8.yaml file or passed via command line:

python src/main.py --conf 0.1 --iou 0.7 --tracker_config configs/botsort.yaml

The system was evaluated on the official Test Set. Below are the obtained results:

• High DetA (0.863): Thanks to the "High Recall" strategy, the system almost never misses players, even in difficult situations.
• Competitive AssA (0.636): Despite the absence of Re-ID, the aggressive use of GMC and tight matching thresholds ensures stable tracking.

• YOLO11: Jocher, G., & Qiu, J. (2024). Ultralytics YOLO11. GitHub.
• SoccerNet: Deliege, A., et al. (2021). SoccerNet-v2: A Dataset and Benchmarks for Holistic Understanding of Broadcast Soccer Videos.
• BoT-SORT: Aharon, N., et al. (2022). BoT-SORT: Robust Associations Multi-Pedestrian Tracking.

Project developed for the Artificial Vision course, A.Y. 2025/2026.