Enhanced temperature measurement using infrared thermography in dynamic environments through an automated robust detection-tracking approach

This work presents a robust approach for temperature measurement in dynamic environments, integrating detection and tracking techniques to enhance accuracy. The proposed method utilises deep learning, particularly convolutional neural networks (CNNs), to detect and track objects of interest within infrared thermography images, eliminating the need for unreliable GPS coordinates. CNNs excel at extracting complex patterns and features from dynamic datasets, enabling effective identification of thermal signatures in varying environmental conditions. The method includes an active learning component to iteratively improve detection and tracking performance, adapting to new data and feedback over time. The proposed system undergoes thorough evaluation, initially using a laboratory prototype to test various configurations, including synthetic false positives and missed detections. The system is then deployed in an industrial facility with a large pipeline system, where an autonomous aerial vehicle performs fully automated inspections, including a possible angle-corrected emissivity handling. A mission planning proposal is also introduced to outline the drone flight execution. The approach addresses several challenges, such as navigation inaccuracies, weather variability, image quality, and processing speed, demonstrating its capacity for accurate temperature measurements even in challenging conditions. Rigorous testing confirms the reliability of the method, highlighting its potential for real-world applications in dynamic industrial environments.