Magnetic confinement fusion reactors are complex devices where a large amount of energy is required to make the fusion reactions happen. In such experimental conditions, the Plasma Facing Components (PFC) are subjected to high heat fluxes. In current tokamaks like Tore Supra, infrared thermographic diagnostics based on image analysis and feedback control are used to measure and monitor the heating of the PFC during plasma operation. The system consists in detecting high increase of the IR luminance signal beyond fixed temperature thresholds for a set of predefined Regions of Interest (ROI). Consequently, this system neither takes into account the thermal objects outside of the ROI, nor the geometric information of the detected thermal object. In this paper, we propose a new vision based approach for the automatic detection of thermal events. This approach is composed of three main tasks: thermal object detection (1), classification (2), and thermal event recognition (3). We present results of our approach for the recognition of one critical thermal event and compare it with the previous system.