Аңдатпа
Objective and reproducible scoring of marksmanship is a long-standing challenge in sport shooting and firearms training, where manual scoring is slow and subjective and dedicated electronic targets are expensive. This paper presents ShootVision, a web-based system that performs real-time detection and accuracy analysis of shots using computer vision. The core of the system is a YOLOv8n (nano) single-stage detector, fine-tuned through transfer learning to perform binary classification of detected objects into two classes, hit and miss. A custom augmentation pipeline expands each source image into six variants through horizontal flipping, rotation with analytical recomputation of bounding-box coordinates, and brightness and contrast variation. To suppress false positives in a live video stream, ShootVision introduces a three-level detection-filtering scheme (confidence thresholding, a custom area-aware non-maximum suppression, and temporal smoothing over a sliding window) together with an original temporal-stability mechanism that confirms a shot only after repeated, spatially consistent detections. Hit accuracy is quantified by the Euclidean distance from each hit to the target centre and mapped to a ten-point score over concentric scoring zones. The system is delivered through a Django web platform that streams annotated video and offers four operating modes (classic, distance, target, and duel), a trilingual interface, and interactive analytics. The deployed system operates at approximately 30 frames per second and demonstrates that lightweight object detection combined with temporal-stability logic provides a low-cost, accessible alternative to specialised scoring hardware.