REAL – TIME VIDEO GAME EVENTS DETECTION USING MACHINE LEARNING

Abstract

The gaming industry has evolved dramatically, with modern video games offering highly immersive and complex environments. Despite this progress, real-time detection and analysis of in-game events remain relatively untapped areas that could significantly enhance player experience and game development. Real-time detection and analysis of in-game events provide valuable insights that can drive dynamic gameplay adjustments, optimize game balance, and personalize content delivery. This thesis addresses the critical need for advanced analytics in gaming by exploring real-time video game event detection using machine learning techniques. By accurately identifying and classifying character statuses and enemy types in "Diablo II: Resurrected," the aim is to empower both developers and players with actionable data that enriches gameplay and supports informed decision-making. To achieve this, we leverage the YOLOv8 deep learning architecture, renowned for its efficiency in real-time object detection, to develop a custom-trained model capable of distinguishing character statuses within the game's dynamic environment. Systematic data collection and precise annotation of in-game screenshots allowed us to create a robust dataset aligned with significant game events. The model underwent rigorous training and fine-tuning, guided by evaluation metrics like precision, recall, and mean average precision (mAP), ensuring high detection accuracy and responsiveness. The results demonstrate the model's effectiveness, validating the suitability of YOLOv8 for complex real-time applications. This research contributes to the fields of machine learning and game AI by advancing real-time detection and classification capabilities in interactive virtual environments, underscoring the pivotal role of analytics in modern gaming.