Automatic Classification and Visualisation of Gas from Infrared Video Data

Sammanfattning: Optical gas imaging denotes the visualisation of gases by means of an infrared camera, which allows operators to quickly, easily, and safely scan a large area and therefore plays a major role in the early detection and repair of gas leaks in various environments within the petrochemical industry such as processing plants and pipelines, but also in production facilities and hospitals. Thereby they help to avert damage to the environment as well as to health and safety of workers or inhabitants of nearby residential areas. The current generation of thermal gas cameras employs a so-called high-sensitivity mode, based on frame differencing, to increase the visibility of gas plumes. However, this method often results in image degradation through loss of orientation, distortion, and additional noise. Taking the increased prevalence and sinking costs for IR gas cameras – entailing an increased number of inexperienced users – into consideration, a more intuitive and user-friendly system to visualise gas constitutes a useful feature for the next generation of IR gas cameras. A system that retains the original infrared video images and highlights the gas cloud, providing the user with a clear and distinct visualisation of gas on the camera’s display, would be one example for such a visualisation system. This thesis discusses the design of such an automatic gas detection and visualisation framework based on machine learning and computer vision methods, where moving objects in video images are detected and classified as gas or non-gas based on appearance and spatiotemporal features. The main goal was to conduct a proof-of-concept study of this method, which included gathering examples for training a classifier as well as implementing the framework and evaluating several feature descriptors – both static and dynamic ones – with regard to their classification performance in gas detection in video images. Depending on the application scenario, the methods evaluated in this study are capable of reliably detecting gas.