Real-Time Close-Proximity Soccer Ball Detection: Model Benchmarking and Inference Optimization
Year: 2026
Type: conference-paper
Abstract: Close-proximity indoor soccer ball detection is more challenging than broadcast-style analysis due to rapid motion, strong scale variation, and frequent motion blur. Using the Prox-iBall dataset and a strictly held-out, view-separated testbench, we benchmark ten modern detectors under both clean and synthetic-blur conditions and evaluate a lightweight temporal-refinement module. RT-DETR-L achieves the highest accuracy on the clean testbench. However, under practical runtime and footprint constraints, YOLOv26s is selected for deployment due to its superior robustness to severe blur and improved recall on fast targets compared to YOLOv10s. All detectors degrade significantly with increasing blur, and stratified analysis shows that small and fast targets remain the primary failure modes. The temporal module is designed for inference throughput—adaptive frame skipping with Kalman propagation targets 100+FPS on edge hardware—and a controlled accuracy ablation reveals a slight recall gain alongside substantially increased false persistence on negative frames. Robust close-proximity ball perception thus requires detector selection by blur robustness under runtime constraints, and pre-deployment validation of temporal propagation’s precision impact.
Institution: Nazarbayev University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7206198509