Performance Evaluation of YOLOv8 for Automatic Bullet Impact Detection and Scoring on Shooting Targets

Closed

Erdiyan Ariyanda, Moechammad Sarosa, Sapto Wibowo, Wahyu Nur Hidayat, Yunia Mulyani Azis, Isa Mahfudi

2026 Proceedings - 2026 International Conference on Current Research in Artificial Intelligence and Data Science, ICCRAIDS 2026 Conference paper Cited by 0 Quartile

Abstract

Automatic scoring on shooting targets is traditionally performed manually, which is time-consuming, subjective, and prone to human error. To address these limitations, this study presents a comprehensive performance evaluation of YOLOv8 for automatic bullet impact detection and scoring on paper-based shooting targets. The proposed system formulates bullet hole detection as a single-class small object detection problem and integrates centroid-based localization with a distance-based scoring algorithm. A dataset consisting of 1000 real shooting target images captured using a low-cost ESP32-CAM at 800 × 600 resolution was prepared and manually annotated. The YOLOv8 model was trained and evaluated using standard object detection metrics, including precision, recall, F1-score, mAP@0.5, and m A P@0.5: 0.95. Experimental results show that YOLOv8 achieves precision of approximately 0.75 and recall of approximately 0.52, with mAP@0.5 around 0.565 and mAP@0.5: 0.95 around 0.24. Confidence threshold analysis indicates an optimal threshold near 0.204, yielding the best F1-score of about 0.61. Localization and scoring evaluations demonstrate that the proposed approach can generate scores consistent with ground truth in most cases, achieving an exact score match of 82.1% and 94.6% accuracy within ± 1 score tolerance. The remaining errors are mainly caused by very small, overlapping, and low-contrast bullet holes, particularly near ring boundaries. © 2026 IEEE.

Affiliations

State Polytechnic of Malang, Department of Electrical Engineering, Malang, Indonesia; STIE Ekuitas Bandung, Bandung, Indonesia