CricketAnalyzer: A Parallel HPC-Accelerated System for Real-Time Cricket Ball Tracking and Speed Detection

Accurate real-time detection of moving object speeds in high-velocity sports like cricket, soccer, tennis, and golf is essential for performance analytics, coaching decisions, and broadcasting improvements. Traditional radar guns are accurate but costly, inflexible, and provide limited analytics in controlled environments only.This paper presents CricketAnalyzer, a vision-based application for real-time cricket ball speed estimation using cameras with varying resolutions and data rates. The system integrates Multi-Region of Interest (Multi-ROI) architecture with parallel Large Vision Transformer (ViT) and Segment Anything Model (SAM) for efficient object detection and motion tracking under dynamic conditions including varying lighting, weather, and camera placement. To enhance small object detection, we introduce a modified YOLOv8 featuring Split Cross-Stage Partial Network with Two-Level Feature Fusion (Split-C2F) optimized for high-precision detection of cricket balls and stumps. Real-time performance is achieved through data-level parallelism using four nodes of the Namal Pak Supercomputing cluster. CricketAnalyzer supports various cameras (RTSP streams, USB, GMSL, GoPro) at 30-240 FPS. Results demonstrate 0.943 precision and 0.904 recall in ball detection, with ±1% speed estimation error and >90% frame-wise tracking consistency. This scalable, cost-effective solution enables continuous ball trajectory computation at high frame rates, offering optimized sports analytics through parallel deep learning inference.