Early fire detection is crucial to minimise damage and save lives. Video surveillance smoke detectors do not suffer from transport delays and can cover large areas. The smoke detection on images is, however, a difficult problem due the variability of smoke density, lighting conditions, background clutter, and unstable patterns. In order to solve this problem, we propose a novel unsupervised object classifier. Single visual features are classified using a model that simultaneously creates a codebook and categorises the smoke using a bag-of-words paradigm based on LDA model. Our algorithm can also tell the amount of smoke present on the image. Multiple image sequences from different cameras are used to show the viability of the proposed approach. Our experiments show that the model generalises well for different cameras, perspectives and scales.