Degree Name

Doctor of Philosophy


School of Computer Science and Software Engineering


To prevent property damage and loss of human lives due to fire accidents, vision-based smoke detection is deemed to be a promising approach to early detection and prediction of fire disasters. Despite advances, challenges still exist for developing robust vision-based smoke detection algorithms, which include reliably quantifying the visual characteris-tics of smoke and overcoming the limitations of devices in real applications. This thesis has established computational models to address these challenges and developed novel methods based on the models for vision-based smoke detection. Specifically, an image formation model for smoke is derived based on atmospheric scattering models. To cap-ture the texture feature of smoke, non-redundant local binary pattern is adopted to deal with the relative intensity between the bright/dark smoke and the background. In addition, algorithms are developed to separate the smoke component from the background and ro-bust features are extracted from the smoke component for reliable detection. Specifically, three models, namely local smoothness, principal component and sparse representation, are proposed to model the smoke component for effective component separation. De-tection performance is significantly improved compared with the traditional video-based methods. For the scenarios where video is not available, novel algorithms are developed for detecting smoke in single images. Specifically, sparse representation based on dual-dictionary modeling is proposed for both the smoke and background components and the sparse coefficients are employed as a feature for smoke detection. The effectiveness of the proposed methods is validated by extensive experiments.