PitchStat

Project Overview

• AI-powered football analysis solution focusing on player tracking, detailed performance statistics, and event identification (passes, shots, and goals).
• Leverages advanced computer vision and machine learning techniques to deliver actionable insights for coaches, analysts, and players.
• Automated tracking and event detection reduced manual analysis time by over 70% for full-match reviews.
• Adaptable for different leagues, match formats, and custom event definitions, ensuring broad applicability.

Challenges

AI-Based Football Analysis Platform

• Player Tracking: State-of-the-art object detection (YOLO) and multi-object tracking algorithms to identify and track each player throughout the match.
• Occlusion & Trajectory Correction: Maintains accurate tracking even when players overlap or move out of view.
• Duration Played: Automatically calculated on-field time for each player based on continuous tracking data.
• Distance Covered: Measured total distance using perspective transformation and trajectory analysis.
• Speed Metrics: Computed both maximum and average speeds for every player, enabling performance benchmarking.
• Pass Detection: Rule-based and ML algorithms to detect passes using spatial and temporal data from player and ball positions.
• Shot and Goal Detection: Classified shots and goals with high accuracy using event-based classifiers and positional data.
• Real-Time Event Logging: Instant annotation of key match events for immediate tactical analysis.

Implementation Approach

• Video Pre-processing (Frame Extraction): Videos converted into individual frames for further processing.
• Frame Annotation (CVAT): Subset of frames (1,000–3,000) annotated using CVAT with bounding boxes around players to create ground truth data.
• Object Detection Model Training (YOLOv8): Annotated frames used to train object detection models for accurate player identification.
• Object Tracking: Multi-object tracking with DeepSORT selected over SimpleSORT, ByteTrack, and OC-SORT for its fine-tuning capabilities.
• Detection Threshold (det_threshold): Minimum confidence score from YOLOv8 for a detection to be considered valid for tracking.
• IoU Threshold (iou_threshold): Used during tracking to determine if a new detection corresponds to an existing tracked object.
• Unique Player ID Assignment: Each player assigned a persistent tracking_id_object for consistent recognition throughout the video.
• Occlusion Management Research: Ongoing research to manage and mitigate occlusion challenges for continuous and robust tracking.

Implementation Timeline

• Requirements & Design (1 week): Analyzed client needs, mapped workflows, and defined key performance indicators for analysis.
• Model Development (2 weeks): Trained and fine-tuned object detection and tracking models; developed event detection logic.
• Data Integration (2 weeks): Processed match videos, integrated player and ball tracking, and computed advanced statistics.
• QA & Deployment (1 week): Validated event detection accuracy, ensured real-time performance, and deployed to production.

Results

• Enhanced Efficiency: Automated tracking reduced manual analysis time by over 70% for full-match reviews.
• Comprehensive Player Insights: Detailed statistics (duration, distance, speed) and event logs for all players.
• High Event Detection Accuracy: F-scores up to 0.92 for pass detection and 0.91 for shot classification.
• Real-Time Feedback: Instant insights for coaches and analysts enabling timely tactical adjustments.
• Scalability: Adaptable for different leagues, match formats, and custom event definitions.

Technical Performance

Operational Efficiency

System Reliability

Business Impact

Technology Stack