A Hybrid Vision-Guided Autonomous Firefighting System with Dynamic Stabilization on a Stewart Platform

This study presents a modular and cost-effective autonomous firefighting system designed to operate reliably under dynamic and unstable environmental conditions. The system architecture is grounded in hardware and software innovations specifically designed to address the limitations of conventional fire detection and targeting methods. To emulate real-world motion disturbances such as tilt, vibration, and oscillation, a six-degree-of-freedom Stewart platform is employed as a testbed. A dual-axis turret mounted on this platform utilizes sensor fusion with Mahony filtering, and PID control to maintain image stability and ensure precise directional alignment despite continuous perturbations. To complement the mechanical architecture, a hybrid fire detection framework is implemented by integrating HSV-based filtering with the YOLOv8 deep learning model, balancing low-latency response with high detection accuracy across varying visual scenarios. Furthermore, a MiDaS-based depth estimation module enhances the system's spatial perception by providing real-time distance information of detected fire sources. Experimental results demonstrate the effectiveness of the proposed system architecture in maintaining detection accuracy and directional stability under simulated real-world disturbances, offering a promising solution for autonomous firefighting applications.